[{"_1":2,"_660":-5,"_661":-5},"loaderData",{"_3":4,"_38":-5,"_39":40,"_509":510,"_658":-5,"_659":-5},"root",{"_5":-5,"_6":-5,"_7":8,"_22":23,"_24":25,"_26":27,"_30":31,"_36":37},"user","shoppingCartId","shoppingCart",{"_9":10,"_11":12,"_13":14,"_15":16,"_17":18,"_19":20,"_21":14},"id","no-cart","count",0,"free",true,"updatedAt","","paymentGateway","STRIPE","cartItems",[],"loading","locale","en","readerTheme","light","constants",{"_28":29},"paypalClientId","AX53Cwz8jR6rVVzzZiZNWWXl0dGagi-GZrjG-S3Ua14E5nrKMn7DxDQ730O4R8LusLPp49ALn6aRVaEy","publishers",{"_32":33,"_34":35},"companies",[],"universities",[],"needsLocaleCookieSync",false,"routes/_store+/_readers+/read_/route","routes/_store+/_readers+/read_.$slug/route",{"_41":42,"_67":68,"_488":489,"_507":37,"_508":-7},"book",{"_43":44,"_45":46,"_47":48,"_49":50,"_53":54,"_56":57,"_65":66},"title","Build Your Own Coding Agent","slug","build-your-own-coding-agent","metaText","Build a coding agent in pure Python. No LangChain or vector DBs. Orchestrate Claude, DeepSeek, or Ollama with raw HTTP requests. Test everything with FakeBrain and pytest. Includes full source code and a capstone where the agent builds a Snake game autonomously.","coverFeatured",{"_51":52},"url","https://d2sofvawe08yqg.cloudfront.net/build-your-own-coding-agent/s_featured?1771968015&1771968015","coverFeatured2x",{"_51":55},"https://d2sofvawe08yqg.cloudfront.net/build-your-own-coding-agent/s_featured2x?1771968015&1771968015","authors",[58],{"_9":59,"_60":61,"_5":62},"Z2lkOi8vbGVhbnB1Yi9TdG9yZTo6UHJvZHVjdEF1dGhvci84MDY3OA","name","J. Owen",{"_63":64},"username","optimalone","showReadLinkString","sample","parts",[69],{"_43":44,"_70":16,"_71":12,"_72":73},"description","fileCount","chapters",[74,102,106,137,163,202,245,249,284,315,343,364,399,403,436,457,484],{"_75":76,"_43":77,"_78":79,"_71":80,"_81":82,"_89":-5},"level",2,"Preface","path","#preface",1,"subsections",[83,90,93,96,99],{"_75":84,"_43":85,"_70":-5,"_78":86,"_71":80,"_87":-5,"_88":-5,"_89":-5},3,"Who This Book Is For","#who-this-book-is-for","quizType","sectionNum","youtubeId",{"_75":84,"_43":91,"_70":-5,"_78":92,"_71":80,"_87":-5,"_88":-5,"_89":-5},"What You Will Build","#what-you-will-build",{"_75":84,"_43":94,"_70":-5,"_78":95,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Testing Approach","#testing-approach",{"_75":84,"_43":97,"_70":-5,"_78":98,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Code Examples","#code-examples",{"_75":84,"_43":100,"_70":-5,"_78":101,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Conventions Used in This Book","#conventions-used-in-this-book",{"_75":76,"_43":103,"_78":104,"_71":80,"_81":105,"_89":-5},"Part I: The Brain","#part-i-the-brain",[],{"_75":76,"_43":107,"_78":108,"_71":80,"_81":109,"_89":-5},"Chapter 1: The Zero Magic Manifesto","#chapter-1-the-zero-magic-manifesto",[110,113,116,119,122,125,128,131,134],{"_75":84,"_43":111,"_70":-5,"_78":112,"_71":80,"_87":-5,"_88":-5,"_89":-5},"What is an Agent, Really?","#what-is-an-agent-really",{"_75":84,"_43":114,"_70":-5,"_78":115,"_71":80,"_87":-5,"_88":-5,"_89":-5},"What We Are Building","#what-we-are-building",{"_75":84,"_43":117,"_70":-5,"_78":118,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Project Setup","#project-setup",{"_75":84,"_43":120,"_70":-5,"_78":121,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The AgentStop Exception","#the-agentstop-exception",{"_75":84,"_43":123,"_70":-5,"_78":124,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Agent Class","#the-agent-class",{"_75":84,"_43":126,"_70":-5,"_78":127,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Defining Success with Tests","#defining-success-with-tests",{"_75":84,"_43":129,"_70":-5,"_78":130,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Main Loop","#the-main-loop",{"_75":84,"_43":132,"_70":-5,"_78":133,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Run It","#run-it",{"_75":84,"_43":135,"_70":-5,"_78":136,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Wrapping Up","#wrapping-up",{"_75":76,"_43":138,"_78":139,"_71":80,"_81":140,"_89":-5},"Chapter 2: The Raw Request","#chapter-2-the-raw-request",[141,144,147,150,153,155,158,161],{"_75":84,"_43":142,"_70":-5,"_78":143,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Get an API Key","#get-an-api-key",{"_75":84,"_43":145,"_70":-5,"_78":146,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Vault (.env)","#the-vault-env",{"_75":84,"_43":148,"_70":-5,"_78":149,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Anatomy of a Request","#the-anatomy-of-a-request",{"_75":84,"_43":151,"_70":-5,"_78":152,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Code","#the-code",{"_75":84,"_43":132,"_70":-5,"_78":154,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#run-it-1",{"_75":84,"_43":156,"_70":-5,"_78":157,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Troubleshooting","#troubleshooting",{"_75":84,"_43":159,"_70":-5,"_78":160,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Cleaning Up","#cleaning-up",{"_75":84,"_43":135,"_70":-5,"_78":162,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-1",{"_75":76,"_43":164,"_78":165,"_71":80,"_81":166,"_89":-5},"Chapter 3: The Infinite Loop","#chapter-3-the-infinite-loop",[167,170,173,176,179,182,185,188,191,194,197,200],{"_75":84,"_43":168,"_70":-5,"_78":169,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Illusion of Memory","#the-illusion-of-memory",{"_75":84,"_43":171,"_70":-5,"_78":172,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Testing Problem","#the-testing-problem",{"_75":84,"_43":174,"_70":-5,"_78":175,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Response Types","#response-types",{"_75":84,"_43":177,"_70":-5,"_78":178,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The FakeBrain Pattern","#the-fakebrain-pattern",{"_75":84,"_43":180,"_70":-5,"_78":181,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Defining Success","#defining-success",{"_75":84,"_43":183,"_70":-5,"_78":184,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Claude Class","#the-claude-class",{"_75":84,"_43":186,"_70":-5,"_78":187,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Agent Class (Updated)","#the-agent-class-updated",{"_75":84,"_43":189,"_70":-5,"_78":190,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Main Loop (Updated)","#the-main-loop-updated",{"_75":84,"_43":192,"_70":-5,"_78":193,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Verify the Tests Pass","#verify-the-tests-pass",{"_75":84,"_43":195,"_70":-5,"_78":196,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Test the Memory","#test-the-memory",{"_75":84,"_43":198,"_70":-5,"_78":199,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Context Window Problem","#the-context-window-problem",{"_75":84,"_43":135,"_70":-5,"_78":201,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-2",{"_75":76,"_43":203,"_78":204,"_71":80,"_81":205,"_89":-5},"Chapter 4: The Universal Adapter","#chapter-4-the-universal-adapter",[206,209,212,215,218,221,224,227,229,232,234,237,240,243],{"_75":84,"_43":207,"_70":-5,"_78":208,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Adapter Pattern","#the-adapter-pattern",{"_75":84,"_43":210,"_70":-5,"_78":211,"_71":80,"_87":-5,"_88":-5,"_89":-5},"HTTP Resilience","#http-resilience",{"_75":84,"_43":213,"_70":-5,"_78":214,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Brain Interface","#the-brain-interface",{"_75":84,"_43":216,"_70":-5,"_78":217,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The FakeBrain (Updated)","#the-fakebrain-updated",{"_75":84,"_43":219,"_70":-5,"_78":220,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Claude Brain (Refactored)","#the-claude-brain-refactored",{"_75":84,"_43":222,"_70":-5,"_78":223,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The DeepSeek Brain","#the-deepseek-brain",{"_75":84,"_43":225,"_70":-5,"_78":226,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The BRAINS Registry","#the-brains-registry",{"_75":84,"_43":186,"_70":-5,"_78":228,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-agent-class-updated-1",{"_75":84,"_43":230,"_70":-5,"_78":231,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Tests for Multi-Brain Support","#tests-for-multi-brain-support",{"_75":84,"_43":189,"_70":-5,"_78":233,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-main-loop-updated-1",{"_75":84,"_43":235,"_70":-5,"_78":236,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Setting Up DeepSeek","#setting-up-deepseek",{"_75":84,"_43":238,"_70":-5,"_78":239,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Try It","#try-it",{"_75":84,"_43":241,"_70":-5,"_78":242,"_71":80,"_87":-5,"_88":-5,"_89":-5},"“We Just Moved Code Around”","#we-just-moved-code-around",{"_75":84,"_43":135,"_70":-5,"_78":244,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-3",{"_75":76,"_43":246,"_78":247,"_71":80,"_81":248,"_89":-5},"Part II: The Hands","#part-ii-the-hands",[],{"_75":76,"_43":250,"_78":251,"_71":80,"_81":252,"_89":-5},"Chapter 5: The Tool Protocol","#chapter-5-the-tool-protocol",[253,256,259,262,265,268,271,274,277,279,282],{"_75":84,"_43":254,"_70":-5,"_78":255,"_71":80,"_87":-5,"_88":-5,"_89":-5},"How Tools Actually Work","#how-tools-actually-work",{"_75":84,"_43":257,"_70":-5,"_78":258,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Defining the Tool Interface","#defining-the-tool-interface",{"_75":84,"_43":260,"_70":-5,"_78":261,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The ReadFile Tool","#the-readfile-tool",{"_75":84,"_43":263,"_70":-5,"_78":264,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The WriteFile Tool","#the-writefile-tool",{"_75":84,"_43":266,"_70":-5,"_78":267,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Tool Helpers","#tool-helpers",{"_75":84,"_43":269,"_70":-5,"_78":270,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Updating the Thought Class","#updating-the-thought-class",{"_75":84,"_43":272,"_70":-5,"_78":273,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Updating the Claude Class","#updating-the-claude-class",{"_75":84,"_43":275,"_70":-5,"_78":276,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Agent Class with Tools","#the-agent-class-with-tools",{"_75":84,"_43":129,"_70":-5,"_78":278,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-main-loop-1",{"_75":84,"_43":280,"_70":-5,"_78":281,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Test It","#test-it",{"_75":84,"_43":135,"_70":-5,"_78":283,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-4",{"_75":76,"_43":285,"_78":286,"_71":80,"_81":287,"_89":-5},"Chapter 6: The Scratchpad (Memory)","#chapter-6-the-scratchpad-memory",[288,291,294,297,300,302,305,308,310,313],{"_75":84,"_43":289,"_70":-5,"_78":290,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The “Zero Magic” Memory","#the-zero-magic-memory",{"_75":84,"_43":292,"_70":-5,"_78":293,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Memory Class","#the-memory-class",{"_75":84,"_43":295,"_70":-5,"_78":296,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The ToolContext Class","#the-toolcontext-class",{"_75":84,"_43":298,"_70":-5,"_78":299,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The SaveMemory Tool","#the-savememory-tool",{"_75":84,"_43":272,"_70":-5,"_78":301,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#updating-the-claude-class-1",{"_75":84,"_43":303,"_70":-5,"_78":304,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Crafting the System Prompt","#crafting-the-system-prompt",{"_75":84,"_43":306,"_70":-5,"_78":307,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Updating the Agent Class","#updating-the-agent-class",{"_75":84,"_43":189,"_70":-5,"_78":309,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-main-loop-updated-2",{"_75":84,"_43":311,"_70":-5,"_78":312,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Testing Persistence","#testing-persistence",{"_75":84,"_43":135,"_70":-5,"_78":314,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-5",{"_75":76,"_43":316,"_78":317,"_71":80,"_81":318,"_89":-5},"Chapter 7: The Safety Harness (Plan Mode)","#chapter-7-the-safety-harness-plan-mode",[319,322,325,328,331,333,335,338,341],{"_75":84,"_43":320,"_70":-5,"_78":321,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Concept","#the-concept",{"_75":84,"_43":323,"_70":-5,"_78":324,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Tests First","#tests-first",{"_75":84,"_43":326,"_70":-5,"_78":327,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Extending ToolContext","#extending-toolcontext",{"_75":84,"_43":329,"_70":-5,"_78":330,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Guarded WriteFile Tool","#the-guarded-writefile-tool",{"_75":84,"_43":186,"_70":-5,"_78":332,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-agent-class-updated-2",{"_75":84,"_43":189,"_70":-5,"_78":334,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#the-main-loop-updated-3",{"_75":84,"_43":336,"_70":-5,"_78":337,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Testing the Harness","#testing-the-harness",{"_75":84,"_43":339,"_70":-5,"_78":340,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Psychology of the “Plan”","#the-psychology-of-the-plan",{"_75":84,"_43":135,"_70":-5,"_78":342,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-6",{"_75":76,"_43":344,"_78":345,"_71":80,"_81":346,"_89":-5},"Chapter 8: The Context Pipeline (Map & Search)","#chapter-8-the-context-pipeline-map--search",[347,350,353,356,359,362],{"_75":84,"_43":348,"_70":-5,"_78":349,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The ListFiles Tool","#the-listfiles-tool",{"_75":84,"_43":351,"_70":-5,"_78":352,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The SearchCodebase Tool","#the-searchcodebase-tool",{"_75":84,"_43":354,"_70":-5,"_78":355,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Update the Tools List","#update-the-tools-list",{"_75":84,"_43":357,"_70":-5,"_78":358,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The “Zoom In” Test","#the-zoom-in-test",{"_75":84,"_43":360,"_70":-5,"_78":361,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Wait, is this RAG?","#wait-is-this-rag",{"_75":84,"_43":135,"_70":-5,"_78":363,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-7",{"_75":76,"_43":365,"_78":366,"_71":80,"_81":367,"_89":-5},"Chapter 9: The Reality Check (Running Code)","#chapter-9-the-reality-check-running-code",[368,371,373,376,379,382,385,388,391,394,397],{"_75":84,"_43":369,"_70":-5,"_78":370,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Feedback Loop","#the-feedback-loop",{"_75":84,"_43":323,"_70":-5,"_78":372,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#tests-first-1",{"_75":84,"_43":374,"_70":-5,"_78":375,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The RunCommand Tool","#the-runcommand-tool",{"_75":84,"_43":377,"_70":-5,"_78":378,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Interactive Trap","#the-interactive-trap",{"_75":84,"_43":380,"_70":-5,"_78":381,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Self-Healing Demo","#the-self-healing-demo",{"_75":84,"_43":383,"_70":-5,"_78":384,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The TDD Workflow","#the-tdd-workflow",{"_75":84,"_43":386,"_70":-5,"_78":387,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Surgical Edit","#the-surgical-edit",{"_75":84,"_43":389,"_70":-5,"_78":390,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Closed Loop","#the-closed-loop",{"_75":84,"_43":392,"_70":-5,"_78":393,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Context Compaction","#context-compaction",{"_75":84,"_43":395,"_70":-5,"_78":396,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Security Considerations","#security-considerations",{"_75":84,"_43":135,"_70":-5,"_78":398,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-8",{"_75":76,"_43":400,"_78":401,"_71":80,"_81":402,"_89":-5},"Part III: The Frontier","#part-iii-the-frontier",[],{"_75":76,"_43":404,"_78":405,"_71":80,"_81":406,"_89":-5},"Chapter 10: Going Dark (Local Models)","#chapter-10-going-dark-local-models",[407,410,413,416,419,422,425,428,431,434],{"_75":84,"_43":408,"_70":-5,"_78":409,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Trade-off","#the-trade-off",{"_75":84,"_43":411,"_70":-5,"_78":412,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Installing Ollama","#installing-ollama",{"_75":84,"_43":414,"_70":-5,"_78":415,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Ollama Brain Class","#the-ollama-brain-class",{"_75":84,"_43":417,"_70":-5,"_78":418,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Running with Ollama","#running-with-ollama",{"_75":84,"_43":420,"_70":-5,"_78":421,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The “Infinite Loop” Experiment","#the-infinite-loop-experiment",{"_75":84,"_43":423,"_70":-5,"_78":424,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Practical Differences","#the-practical-differences",{"_75":84,"_43":426,"_70":-5,"_78":427,"_71":80,"_87":-5,"_88":-5,"_89":-5},"The Hybrid Workflow","#the-hybrid-workflow",{"_75":84,"_43":429,"_70":-5,"_78":430,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Model Selection","#model-selection",{"_75":84,"_43":432,"_70":-5,"_78":433,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Troubleshooting Ollama","#troubleshooting-ollama",{"_75":84,"_43":135,"_70":-5,"_78":435,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-9",{"_75":76,"_43":437,"_78":438,"_71":80,"_81":439,"_89":-5},"Chapter 11: The Extension (Web Search)","#chapter-11-the-extension-web-search",[440,443,446,449,452,455],{"_75":84,"_43":441,"_70":-5,"_78":442,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 1: The Meta-Prompt","#step-1-the-meta-prompt",{"_75":84,"_43":444,"_70":-5,"_78":445,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 2: The Surgery","#step-2-the-surgery",{"_75":84,"_43":447,"_70":-5,"_78":448,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 3: The Reference Implementation","#step-3-the-reference-implementation",{"_75":84,"_43":450,"_70":-5,"_78":451,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 4: The Tests","#step-4-the-tests",{"_75":84,"_43":453,"_70":-5,"_78":454,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Self-Modification","#self-modification",{"_75":84,"_43":135,"_70":-5,"_78":456,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-10",{"_75":76,"_43":458,"_78":459,"_71":80,"_81":460,"_89":-5},"Chapter 12: The Capstone (Building a Game)","#chapter-12-the-capstone-building-a-game",[461,464,467,470,473,476,479,481],{"_75":84,"_43":462,"_70":-5,"_78":463,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 1: Preparation","#step-1-preparation",{"_75":84,"_43":465,"_70":-5,"_78":466,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 2: The Architect (Plan Mode)","#step-2-the-architect-plan-mode",{"_75":84,"_43":468,"_70":-5,"_78":469,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 3: The Builder (Act Mode)","#step-3-the-builder-act-mode",{"_75":84,"_43":471,"_70":-5,"_78":472,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 4: The Reality Check","#step-4-the-reality-check",{"_75":84,"_43":474,"_70":-5,"_78":475,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Step 5: The Pivot (Feature Creep)","#step-5-the-pivot-feature-creep",{"_75":84,"_43":477,"_70":-5,"_78":478,"_71":80,"_87":-5,"_88":-5,"_89":-5},"What Goes Wrong","#what-goes-wrong",{"_75":84,"_43":135,"_70":-5,"_78":480,"_71":80,"_87":-5,"_88":-5,"_89":-5},"#wrapping-up-11",{"_75":84,"_43":482,"_70":-5,"_78":483,"_71":80,"_87":-5,"_88":-5,"_89":-5},"Epilogue","#epilogue",{"_75":76,"_43":485,"_78":486,"_71":80,"_81":487,"_89":-5},"Acknowledgments","#acknowledgments",[],"toc",[490,83,90,93,96,99,491,492,110,113,116,119,122,125,128,131,134,493,141,144,147,150,153,155,158,161,494,167,170,173,176,179,182,185,188,191,194,197,200,495,206,209,212,215,218,221,224,227,229,232,234,237,240,243,496,497,253,256,259,262,265,268,271,274,277,279,282,498,288,291,294,297,300,302,305,308,310,313,499,319,322,325,328,331,333,335,338,341,500,347,350,353,356,359,362,501,368,371,373,376,379,382,385,388,391,394,397,502,503,407,410,413,416,419,422,425,428,431,434,504,440,443,446,449,452,455,505,461,464,467,470,473,476,479,481,506],{"_75":76,"_43":77,"_70":-5,"_78":79,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":103,"_70":-5,"_78":104,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":107,"_70":-5,"_78":108,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":138,"_70":-5,"_78":139,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":164,"_70":-5,"_78":165,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":203,"_70":-5,"_78":204,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":246,"_70":-5,"_78":247,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":250,"_70":-5,"_78":251,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":285,"_70":-5,"_78":286,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":316,"_70":-5,"_78":317,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":344,"_70":-5,"_78":345,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":365,"_70":-5,"_78":366,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":400,"_70":-5,"_78":401,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":404,"_70":-5,"_78":405,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":437,"_70":-5,"_78":438,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":458,"_70":-5,"_78":459,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":485,"_70":-5,"_78":486,"_71":80,"_87":-5,"_88":-5,"_89":-5},"isUnavailableOnline","unavailableMessage","routes/_store+/_readers+/read_.$slug.$chapterId/route",{"_43":44,"_511":512,"_41":513,"_655":573,"_656":657},"content","

Part II: The Hands

https://leanpub.com/build-your-own-coding-agent",{"_514":515,"_43":44,"_47":48,"_45":46,"_65":66,"_516":517,"_654":512},"previewContent","

\nPreface\n
- \nWho This Book Is For\n
- \nWhat You Will Build\n
- \nTesting Approach\n
- \nCode Examples\n
- \nConventions Used in This Book\n
\n
\nPart I: The Brain\n
\nChapter 1: The Zero Magic Manifesto\n
- \nWhat is an Agent, Really?\n
- \nWhat We Are Building\n
- \nProject Setup\n
- \nThe AgentStop Exception\n
- \nThe Agent Class\n
- \nDefining Success with Tests\n
- \nThe Main Loop\n
- \nRun It\n
- \nWrapping Up\n
\n
\nChapter 2: The Raw Request\n
- \nGet an API Key\n
- \nThe Vault (.env)\n
- \nThe Anatomy of a Request\n
- \nThe Code\n
- \nRun It\n
- \nTroubleshooting\n
- \nCleaning Up\n
- \nWrapping Up\n
\n
\nChapter 3: The Infinite Loop\n
- \nThe Illusion of Memory\n
- \nThe Testing Problem\n
- \nResponse Types\n
- \nThe FakeBrain Pattern\n
- \nDefining Success\n
- \nThe Claude Class\n
- \nThe Agent Class (Updated)\n
- \nThe Main Loop (Updated)\n
- \nVerify the Tests Pass\n
- \nTest the Memory\n
- \nThe Context Window Problem\n
- \nWrapping Up\n
\n
\nChapter 4: The Universal Adapter\n
- \nThe Adapter Pattern\n
- \nHTTP Resilience\n
- \nThe Brain Interface\n
- \nThe FakeBrain (Updated)\n
- \nThe Claude Brain (Refactored)\n
- \nThe DeepSeek Brain\n
- \nThe BRAINS Registry\n
- \nThe Agent Class (Updated)\n
- \nTests for Multi-Brain Support\n
- \nThe Main Loop (Updated)\n
- \nSetting Up DeepSeek\n
- \nTry It\n
- \n“We Just Moved Code Around”\n
- \nWrapping Up\n
\n
\nPart II: The Hands\n
\nChapter 5: The Tool Protocol\n
- \nHow Tools Actually Work\n
- \nDefining the Tool Interface\n
- \nThe ReadFile Tool\n
- \nThe WriteFile Tool\n
- \nTool Helpers\n
- \nUpdating the Thought Class\n
- \nUpdating the Claude Class\n
- \nThe Agent Class with Tools\n
- \nThe Main Loop\n
- \nTest It\n
- \nWrapping Up\n
\n
\nChapter 6: The Scratchpad (Memory)\n
- \nThe “Zero Magic” Memory\n
- \nThe Memory Class\n
- \nThe ToolContext Class\n
- \nThe SaveMemory Tool\n
- \nUpdating the Claude Class\n
- \nCrafting the System Prompt\n
- \nUpdating the Agent Class\n
- \nThe Main Loop (Updated)\n
- \nTesting Persistence\n
- \nWrapping Up\n
\n
\nChapter 7: The Safety Harness (Plan Mode)\n
- \nThe Concept\n
- \nTests First\n
- \nExtending ToolContext\n
- \nThe Guarded WriteFile Tool\n
- \nThe Agent Class (Updated)\n
- \nThe Main Loop (Updated)\n
- \nTesting the Harness\n
- \nThe Psychology of the “Plan”\n
- \nWrapping Up\n
\n
\nChapter 8: The Context Pipeline (Map & Search)\n
- \nThe ListFiles Tool\n
- \nThe SearchCodebase Tool\n
- \nUpdate the Tools List\n
- \nThe “Zoom In” Test\n
- \nWait, is this RAG?\n
- \nWrapping Up\n
\n
\nChapter 9: The Reality Check (Running Code)\n
- \nThe Feedback Loop\n
- \nTests First\n
- \nThe RunCommand Tool\n
- \nThe Interactive Trap\n
- \nThe Self-Healing Demo\n
- \nThe TDD Workflow\n
- \nThe Surgical Edit\n
- \nThe Closed Loop\n
- \nContext Compaction\n
- \nSecurity Considerations\n
- \nWrapping Up\n
\n
\nPart III: The Frontier\n
\nChapter 10: Going Dark (Local Models)\n
- \nThe Trade-off\n
- \nInstalling Ollama\n
- \nThe Ollama Brain Class\n
- \nRunning with Ollama\n
- \nThe “Infinite Loop” Experiment\n
- \nThe Practical Differences\n
- \nThe Hybrid Workflow\n
- \nModel Selection\n
- \nTroubleshooting Ollama\n
- \nWrapping Up\n
\n
\nChapter 11: The Extension (Web Search)\n
- \nStep 1: The Meta-Prompt\n
- \nStep 2: The Surgery\n
- \nStep 3: The Reference Implementation\n
- \nStep 4: The Tests\n
- \nSelf-Modification\n
- \nWrapping Up\n
\n
\nChapter 12: The Capstone (Building a Game)\n
- \nStep 1: Preparation\n
- \nStep 2: The Architect (Plan Mode)\n
- \nStep 3: The Builder (Act Mode)\n
- \nStep 4: The Reality Check\n
- \nStep 5: The Pivot (Feature Creep)\n
- \nWhat Goes Wrong\n
- \nWrapping Up\n
- \nEpilogue\n
\n
\nAcknowledgments\n

Preface

Who This Book Is For

You are a software engineer who is skeptical of AI hype.

You have seen the demos. You have tried the frameworks. You have watched your LangChain app hallucinate its way into deleting a production database. And you think: “There has to be a better way.”

There is. This book is for developers who want to understand what is actually happening when an AI agent runs. Not the marketing diagrams. Not the “Reasoning Engine” abstractions. The actual HTTP requests. The actual while loop.

If you can read Python and have built a web app or CLI tool before, you have everything you need.

What You Will Build

Nanocode is a coding agent that runs in your terminal. By the end of this book, it will:

\nRead and write files in your codebase\n
\nExecute shell commands\n
\nSearch code using pure Python\n
\nRemember context across sessions\n
\nAsk for permission before dangerous operations\n
\nSearch the web for documentation and answers\n

You will build it from scratch using only requests, subprocess, and pytest. No LangChain. No vector databases. No “orchestration frameworks.” Just Python you can debug with print().

The one exception: Chapter 11 adds ddgs for web search—a single lightweight dependency.

Testing Approach

This book uses a “test-alongside” approach. For each feature:

\nWe introduce the concept—why we need this\n
\nWe show the test first, so you know what success looks like\n
\nWe implement the code to make the test pass\n
\nWe verify with pytest\n

This teaches Test-Driven Development thinking without the full red-green-refactor ceremony in print. It also solves a critical problem: you cannot test an LLM-powered application by actually calling the LLM. API calls are slow, expensive, and non-deterministic.

Starting in Chapter 3, we introduce the FakeBrain pattern—a test double that returns predictable responses. This lets you run your entire test suite without making a single API call or spending a single cent.

Code Examples

This book follows a “Code First” approach. Each chapter builds on the previous one, and every code example is extracted from working files.

To get the code:

\nGitHub: Clone or download from GitHub.¹\n
\nLeanpub: The complete source code is also included in the downloadable resources with your purchase.\n

The code is organized by chapter (ch01/, ch02/, etc.). Each folder contains:

\nnanocode.py — The complete, runnable agent for that chapter\n
\ntest_nanocode.py — Tests that verify the code works without API calls\n

You can copy any chapter folder and pick up from there. Run pytest to verify your code matches the expected behavior.

Conventions Used in This Book

Throughout the book, you will see three types of callouts:

\n\n\n\n\n

$\"An$

Warning: Security or safety risks. Pay attention to these—ignoring them can lead to deleted files or leaked API keys.

\n\n\n\n

$\"An$

Aside: Deep dives and tangents. Useful context, but you can skip them on a first read without losing the thread.

Code walkthroughs follow a consistent pattern: first the context (why we need this code), then the code itself, then a line-by-line explanation of the important parts.

Time to write some code.

\n\n

https://github.com/optimalone/build-your-own-coding-agent↩︎

Part I: The Brain

In these opening chapters, you’ll build the skeleton of an agent: a while loop that talks to Claude via raw HTTP. By Chapter 4, you’ll support multiple LLM providers through the Adapter Pattern—and understand that an “AI agent” is just a loop, a brain, and a list of messages.

Chapter 1: The Zero Magic Manifesto

If you have tried to build an AI application in the last two years, you have likely felt “Framework Fatigue.”

You install a popular library. You import a ReasoningEngine. You call .run(). It works like magic for the “Hello World” example. But the moment you try to do something real—like editing a specific line in a Python file without deleting the imports—it breaks.

And because you used a framework, you can’t fix it. You are stuck digging through layers of abstract classes, factory patterns, and “Chains” trying to find the one prompt that is causing the hallucination.

We are not going to do that here.

This book is a rebellion against “Magic.” We are going to take the “Zero Magic” approach: build a production-grade coding agent called Nanocode in pure Python. No LangChain, no AutoGPT, no Pydantic.

Why? Because an autonomous agent is not magic. It is just a while loop.

What is an Agent, Really?

Strip away the venture capital marketing, and an “Agent” is just a thermostat.

A thermostat reads the temperature (input), compares it to the target (decision), and turns on the heater (action). Then it waits and repeats. That’s it. An AI agent does the same thing, just with text instead of temperature.

\"The — Figure 1. The Agent Loop: User input flows through a while loop where Input (Sensor) feeds the Brain/LLM (Controller), which triggers Tools (Actuator), cycling back to Input until the Brain outputs a Response.

More specifically, an agent has four parts. The Brain is the LLM—a stateless function where you send text and it returns text. It calls Tools—functions like Read File and Run Command—to interact with the outside world. All of this sits inside a Loop (a while True) that keeps cycling until the task is done, with Memory—just a Python list—accumulating the conversation history along the way. (The list dies when the program ends; we’ll add persistent storage in Chapter 6.)

If you can write a while loop, you can build an agent.

By building it from scratch, you will have something the framework users don’t: control. When our agent gets stuck in a loop, you will know exactly which line of code caused it. When the API bill gets too high, you will see exactly where the tokens are leaking.

What We Are Building

Nanocode is a CLI tool that runs in your terminal. You talk to it like a colleague. It reads your files. It runs your commands. It edits your code.

By the end of this book, you will have wired it up to Claude Sonnet 4.6 (or DeepSeek, or a local model via Ollama). You’ll give it hands—tools to read files, write files, and run shell commands—and eyes to search your codebase. And you’ll build a safety harness so it can’t accidentally rm -rf /.

Project Setup

\n\n

1. Initialize the Project

1 mkdir nanocode\n2 cd nanocode\n3 git init\n

\n\n

2. Create a Virtual Environment

Never install AI tools globally. They conflict with system packages.

1 # Mac/Linux\n2 python3 -m venv venv\n3 source venv/bin/activate\n4 \n5 # Windows\n6 python -m venv venv\n7 venv\\Scripts\\activate\n

\n\n

3. Install Dependencies

We only need three libraries:

\nrequests — To talk to the LLM APIs.\n
\npython-dotenv — To load API keys from a .env file.\n
\npytest — To test our code without making API calls.\n

Create requirements.txt:

1 requests\n2 python-dotenv\n3 pytest\n

\n\n

Install:

1 pip install -r requirements.txt\n

\n\n

4. Secure Your Keys

\n\n\n\n

$\"An$

Warning: If you push your API key to GitHub, bots will scrape it and drain your account within minutes.

Create .gitignore:

1 .env\n2 __pycache__/\n3 venv/\n4 .DS_Store\n5 .nanocode/\n

\n\n

The AgentStop Exception

Before we write the event loop, we need a clean exit mechanism. An exception beats break statements scattered across the code.

The Context: Exceptions are not just for errors; they are also a control flow mechanism. When the user types /q, we raise AgentStop. The main loop catches it and exits cleanly.

The Code:

1 # --- Exceptions ---\n2 \n3 class AgentStop(Exception):\n4     \"\"\"Raised when the agent should stop processing.\"\"\"\n5     pass\n

\n\n

This goes at the top of nanocode.py. It’s a marker exception—no logic, just a signal.

The Agent Class

Now the core abstraction: the Agent class. It holds state and logic in one place, which makes it easy to test.

The Context: We could put all the logic in main(). But then we’d have to mock input() and print() to test it. By extracting the logic into Agent.handle_input(), we can test it directly.

The Code:

10 class Agent:\n11     \"\"\"A coding agent that processes user input.\"\"\"\n12 \n13     def __init__(self):\n14         pass\n15 \n16     def handle_input(self, user_input):\n17         \"\"\"Handle user input. Returns output string, raises AgentStop to quit.\"\"\"\n18         if user_input.strip() == \"/q\":\n19             raise AgentStop()\n20 \n21         if not user_input.strip():\n22             return \"\"\n23 \n24         return f\"You said: {user_input}\\n(Agent not yet connected)\"\n

\n\n

The Walkthrough:

\nLines 13-14: Empty constructor for now. We’ll add brain and tools in later chapters.\n
\nLines 18-19: Check for the /q quit command. Raise AgentStop instead of returning a special value.\n
\nLines 21-22: Skip empty input. Return empty string (no output to display).\n
\nLine 24: Echo the input back. This is a placeholder—later, we’ll send this to the Brain.\n

Defining Success with Tests

Before we write the main loop, we need tests.

Create test_nanocode.py:

 1 import pytest\n 2 from nanocode import Agent, AgentStop\n 3 \n 4 \n 5 def test_handle_input_returns_string():\n 6     \"\"\"Verify handle_input returns a string for normal input.\"\"\"\n 7     agent = Agent()\n 8     result = agent.handle_input(\"hello\")\n 9     assert isinstance(result, str)\n10     assert \"hello\" in result\n11 \n12 \n13 def test_empty_input_returns_empty_string():\n14     \"\"\"Verify empty/whitespace input returns empty string.\"\"\"\n15     agent = Agent()\n16     assert agent.handle_input(\"\") == \"\"\n17     assert agent.handle_input(\"   \") == \"\"\n18     assert agent.handle_input(\"\\n\") == \"\"\n19 \n20 \n21 def test_quit_command_raises_agent_stop():\n22     \"\"\"Verify /q raises AgentStop exception.\"\"\"\n23     agent = Agent()\n24     with pytest.raises(AgentStop):\n25         agent.handle_input(\"/q\")\n26 \n27 \n28 def test_quit_command_with_whitespace():\n29     \"\"\"Verify /q works with surrounding whitespace.\"\"\"\n30     agent = Agent()\n31     with pytest.raises(AgentStop):\n32         agent.handle_input(\"  /q  \")\n

\n\n

Run the tests:

1 pytest test_nanocode.py -v\n

\n\n

1 test_nanocode.py::test_handle_input_returns_string PASSED\n2 test_nanocode.py::test_empty_input_returns_empty_string PASSED\n3 test_nanocode.py::test_quit_command_raises_agent_stop PASSED\n4 test_nanocode.py::test_quit_command_with_whitespace PASSED\n

\n\n

All green. Our agent handles the basic cases correctly.

\n\n\n\n

$\"An$

Aside: Why pytest? It discovers functions starting with test_ and runs them. No boilerplate, no classes required. The test code itself is plain Python—no magic.

The Main Loop

Now the thin I/O wrapper that connects the agent to the terminal:

29 def main():\n30     agent = Agent()\n31     print(\"⚡ Nanocode v0.1 initialized.\")\n32     print(\"Type '/q' to quit.\")\n33 \n34     while True:\n35         try:\n36             user_input = input(\"\\n❯ \")\n37             output = agent.handle_input(user_input)\n38             if output:\n39                 print(output)\n40 \n41         except (AgentStop, KeyboardInterrupt):\n42             print(\"\\nExiting...\")\n43             break\n44 \n45 \n46 if __name__ == \"__main__\":\n47     main()\n

\n\n

The Walkthrough:

\nLines 30-32: Create the agent and print startup messages.\n
\nLine 36: input() blocks and waits for the user to type something.\n
\nLines 37-39: Call handle_input() and print any output.\n
\nLines 41-43: Catch AgentStop (from /q) or KeyboardInterrupt (from Ctrl+C) and exit cleanly.\n

\n\n\n\n

$\"An$

Aside: Python’s input() reads one line at a time. All prompts in this book are single-line. This keeps the code simple—production agents use richer input methods like readline or full TUIs.

Notice the separation: Agent.handle_input() contains all the logic. main() is just I/O glue. This makes the agent testable without mocking stdin/stdout.

Run It

1 python nanocode.py\n

\n\n

You should see:

 1 ⚡ Nanocode v0.1 initialized.\n 2 Type '/q' to quit.\n 3 \n 4 ❯ hello\n 5 You said: hello\n 6 (Agent not yet connected)\n 7 \n 8 ❯ /q\n 9 \n10 Exiting...\n

\n\n

This is the chassis. The engine comes next.

Wrapping Up

That’s the chassis: an Agent class, a handle_input() method, a while True loop. It doesn’t do anything useful yet—but everything we build from here plugs into this skeleton. The tests make sure we don’t break what already works as we go.

Chapter 2: The Raw Request

Most tutorials will tell you to pip install anthropic. We are not going to do that.

SDKs hide the truth. They add layers of abstraction that make “Hello World” easy but make debugging “Error 400” a nightmare. When you learn the SDK, you learn the SDK. When you learn the raw HTTP call, you learn the protocol underneath every SDK.

We are going to send a message to Claude using nothing but the requests library. Claude is Anthropic’s flagship LLM—one of the most capable models for coding tasks.

Get an API Key

To talk to Claude, you need an API Key. This is a long string of characters that acts as your credit card.

\nGo to the Anthropic Console.¹\n
\nSign up and add a payment method (minimum $5 credit).\n
\nCreate a new API Key and name it nanocode.\n
\nCopy the key (it starts with sk-ant-...).\n

\n\n\n\n

$\"An$

Warning: Treat this key like a password. Anyone with it can spend your money.

The Vault (.env)

We need a safe place to store this key. We never put keys directly in code.

Create a file named .env in your project root:

1 touch .env\n

\n\n

Open it and paste your key:

1 ANTHROPIC_API_KEY=sk-ant-api03-...\n

\n\n

We installed python-dotenv in Chapter 1 for exactly this purpose—it reads .env and loads the values into os.environ.

The Anatomy of a Request

To talk to an LLM, we send an HTTP POST request to:

https://api.anthropic.com/v1/messages

This request needs three things: authentication in the headers (your API key), configuration in the body (which model, how many tokens), and the message itself.

The Headers

Anthropic requires three headers:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

Header	Value	Purpose
`x-api-key`	Your secret key	Authentication
`anthropic-version`	`2023-06-01`	API version
`content-type`	`application/json`	Format

The Payload

The “Messages API” expects a list of message dictionaries:

1 \"messages\": [\n2     {\"role\": \"user\", \"content\": \"Hello, world!\"}\n3 ]\n

\n\n

Each message has a role (either \"user\" or \"assistant\") and content (the text).

The Code

Create a file called test_api.py. This is a “smoke test” to prove our connection works. We will delete it later.

The Context: We are writing linear, procedural code. No functions, no classes. We want to see the bare metal.

The Code:

 1 import os\n 2 import requests\n 3 import json\n 4 from dotenv import load_dotenv\n 5 \n 6 # 1. Load the vault\n 7 load_dotenv()\n 8 api_key = os.getenv(\"ANTHROPIC_API_KEY\")\n 9 \n10 # Basic check so we don't crash with a confusing \"NoneType\" error later\n11 if not api_key:\n12     print(\"Error: ANTHROPIC_API_KEY not found in .env\")\n13     exit(1)\n14 \n15 # 2. Define the target\n16 url = \"https://api.anthropic.com/v1/messages\"\n17 \n18 # 3. Authenticate\n19 headers = {\n20     \"x-api-key\": api_key,\n21     \"anthropic-version\": \"2023-06-01\",\n22     \"content-type\": \"application/json\"\n23 }\n24 \n25 # 4. Construct the payload\n26 payload = {\n27     \"model\": \"claude-sonnet-4-6\",\n28     \"max_tokens\": 4096,\n29     \"messages\": [\n30         {\"role\": \"user\", \"content\": \"Hello, are you ready to code?\"}\n31     ]\n32 }\n33 \n34 # 5. Fire! (No safety net)\n35 print(\"📡 Sending request to Claude...\")\n36 response = requests.post(url, headers=headers, json=payload, timeout=120)\n37 \n38 # 6. Inspect the raw result\n39 print(f\"Status: {response.status_code}\")\n40 \n41 if response.status_code == 200:\n42     # Success: Print the beautiful JSON\n43     print(\"Response:\")\n44     print(json.dumps(response.json(), indent=2))\n45 else:\n46     # Failure: Print the ugly raw text so we can debug\n47     print(\"Error:\", response.text)\n

\n\n

The Walkthrough:

\nLine 7: load_dotenv() finds the .env file and loads variables into os.environ.\n
\nLine 8: We grab the API key. Never hardcode this.\n
\nLines 11-13: Basic sanity check. Without this, a missing key causes a confusing NoneType error in the headers dict.\n
\nLine 21: The anthropic-version header is mandatory. Omit it, and the API rejects you.\n
\nLine 27: claude-sonnet-4-6 specifies which model we want.\n
\nLine 28: max_tokens is required. It caps the response length and prevents runaway costs.\n
\nLine 36: We fire the request with a 2-minute timeout. No try/except—if the network is down, let Python crash. You need to see where it fails.\n
\nLines 41-47: Check the status code. 200 means success (pretty-print the JSON). Anything else means we print the raw error text for debugging.\n

Run It

1 python test_api.py\n

\n\n

If everything works, you should see:

Status: 200\nResponse:\n{\n  \"id\": \"msg_01...\",\n  \"type\": \"message\",\n  \"role\": \"assistant\",\n  \"content\": [\n    {\n      \"type\": \"text\",\n      \"text\": \"Hello! Yes, I'm ready to code...\"\n    }\n  ],\n  \"stop_reason\": \"end_turn\",\n  \"usage\": {\n    \"input_tokens\": 15,\n    \"output_tokens\": 81\n  }\n}\n

\n\n

Troubleshooting

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

Error	Cause	Fix
401 Unauthorized	Bad API key	Check `.env` is loading. Print `os.environ.get(\"ANTHROPIC_API_KEY\")` to verify.
400 Bad Request	Malformed JSON	Did you forget `max_tokens`? Is `messages` a list?
429 Rate Limit	Too many requests or no credits	Wait, or add credits to your account.

Cleaning Up

We proved we can talk to the brain. Delete test_api.py—we’re done with it.

\n\n\n\n

$\"An$

Aside: This test_api.py is throwaway code—a one-time smoke test. Real automated tests (with FakeBrain and pytest) come in Chapter 3. You should always delete this file after verifying your API connection works.

\n\n\n\n\n

$\"An$

Aside: To monitor your spending, check the Usage tab in the Anthropic Console. As of early 2026, a typical coding session with 20-30 exchanges costs $0.10-$0.50 with Claude Sonnet. The usage field at the bottom of the response JSON shows exact token counts—you could log these to track costs programmatically.

Wrapping Up

That’s a raw API call: headers, JSON payload, response parsing. No abstraction between you and the wire. When something breaks (and it will), you’ll know exactly which layer failed because there’s only one layer.

One problem: Claude has total amnesia. Every request is a blank slate. We’re going to fake memory by replaying the entire conversation history on every turn.

\n\n

https://console.anthropic.com/settings/keys↩︎

Chapter 3: The Infinite Loop

We have a problem.

Run the script from Chapter 2 twice. Say “My name is Alice.” Claude says hello. Run it again and ask “What is my name?” Claude says “I don’t know.”

This is because LLMs are stateless. They have total amnesia. Every request is the first time you have ever met.

To build an agent, we need to fix this by creating artificial memory.

The Illusion of Memory

“Memory” in an LLM isn’t a hard drive. It’s a log file.

When you chat with ChatGPT, it doesn’t “remember” what you said 5 minutes ago. Behind the scenes, the code sends the entire conversation history back to the model with every new message.

The model sees the full transcript every time. That’s the trick.

We are going to implement this context loop manually. But first, we need to make our code testable.

The Testing Problem

Here’s a hard truth: you cannot test an LLM-powered application by actually calling the LLM.

API calls are slow (2-10 seconds each), expensive (real money per call), and non-deterministic (you might get a different response every time). Imagine running a test suite that costs $5 and takes 20 minutes. You’d never run it.

The solution is dependency injection. Instead of hardcoding the API call inside our agent, we pass in a “brain” object. In production, the brain is Claude. In tests, the brain is a fake that returns predictable responses.

We’ll establish this pattern now, before writing any more production code.

Response Types

Before we build the brain, we need to define what it returns. Claude’s API sends back complex JSON with multiple content blocks. We need simple Python objects to work with.

The Context: Claude can return text, tool calls, or both in a single response. We need data classes to represent these possibilities.

The Code:

17 class ToolCall:\n18     \"\"\"A tool invocation request from the brain.\"\"\"\n19 \n20     def __init__(self, id, name, args):\n21         self.id = id\n22         self.name = name\n23         self.args = args  # dict\n

\n\n

ToolCall represents the brain asking us to execute a tool. The id is a unique identifier for tracking (Claude needs it when we report results back). The name is which tool to run. The args is a dictionary of parameters.

We won’t use ToolCall yet—the brain can’t call tools—but we define it now because it’s part of the Thought response type. When we add tools, Claude will return these when it wants to read a file or execute a command.

26 class Thought:\n27     \"\"\"Standardized response from any Brain.\"\"\"\n28 \n29     def __init__(self, text=None, tool_calls=None):\n30         self.text = text  # str or None\n31         self.tool_calls = tool_calls or []  # list of ToolCall\n

\n\n

A Thought is what the brain returns after thinking. It might have text, tool calls, both, or neither. This abstraction will let us swap Claude for DeepSeek later without changing any other code.

The FakeBrain Pattern

Now we can build a fake brain for testing.

The Context: We need a brain that returns predictable responses, tracks how many times it was called, and records what conversation it received.

The Code:

class FakeBrain:\n    \"\"\"Fake brain for testing - returns predictable responses.\"\"\"\n\n    def __init__(self, responses=None):\n        self.responses = responses or [Thought(text=\"Fake response\")]\n        self.call_count = 0\n        self.last_conversation = None\n\n    def think(self, conversation):\n        self.last_conversation = list(conversation)  # Store a copy\n        if self.call_count < len(self.responses):\n            response = self.responses[self.call_count]\n            self.call_count += 1\n            return response\n        return Thought(text=\"No more responses\")\n

\n\n

This goes in test_nanocode.py, not the production code. Notice that FakeBrain has the same interface as our real brain will—a think() method that takes a conversation and returns a Thought.

\n\n\n\n

$\"An$

Aside: This pattern—replacing a real dependency with a predictable fake for testing—is called dependency injection. Martin Fowler’s article “Mocks Aren’t Stubs”¹ explains the variations (fakes, stubs, mocks, spies). For LLM testing, a simple fake with canned responses is usually all you need.

Defining Success

Before writing the production code, let’s define what success looks like. These tests will guide our implementation.

Test 1: The brain returns a response

1 def test_handle_input_returns_brain_response():\n2     \"\"\"Verify handle_input returns the brain's response text.\"\"\"\n3     brain = FakeBrain(responses=[Thought(text=\"Hello from brain!\")])\n4     agent = Agent(brain=brain)\n5     result = agent.handle_input(\"hi\")\n6     assert result == \"Hello from brain!\"\n

\n\n

Notice we pass brain=brain to the Agent. This is dependency injection in action.

Test 2: Conversation accumulates

 1 def test_conversation_accumulates():\n 2     \"\"\"Verify conversation list grows with each interaction.\"\"\"\n 3     brain = FakeBrain(responses=[\n 4         Thought(text=\"Response 1\"),\n 5         Thought(text=\"Response 2\")\n 6     ])\n 7     agent = Agent(brain=brain)\n 8 \n 9     agent.handle_input(\"First message\")\n10     assert len(agent.conversation) == 2  # user + assistant\n11 \n12     agent.handle_input(\"Second message\")\n13     assert len(agent.conversation) == 4  # 2 users + 2 assistants\n

\n\n

After each exchange, the conversation should have both the user message and the assistant response.

Test 3: Correct message structure

 1 def test_conversation_contains_correct_roles():\n 2     \"\"\"Verify conversation has correct role alternation.\"\"\"\n 3     brain = FakeBrain(responses=[Thought(text=\"AI response\")])\n 4     agent = Agent(brain=brain)\n 5 \n 6     agent.handle_input(\"User message\")\n 7 \n 8     assert agent.conversation[0][\"role\"] == \"user\"\n 9     assert agent.conversation[0][\"content\"] == \"User message\"\n10     assert agent.conversation[1][\"role\"] == \"assistant\"\n11     assert agent.conversation[1][\"content\"] == \"AI response\"\n

\n\n

The messages must have the exact format Claude expects: {\"role\": \"user\", \"content\": \"...\"}.

Test 4: Brain receives the conversation

 1 def test_brain_receives_conversation():\n 2     \"\"\"Verify brain.think is called with the conversation list.\"\"\"\n 3     brain = FakeBrain()\n 4     agent = Agent(brain=brain)\n 5 \n 6     agent.handle_input(\"Test message\")\n 7 \n 8     assert brain.last_conversation is not None\n 9     assert len(brain.last_conversation) == 1\n10     assert brain.last_conversation[0][\"content\"] == \"Test message\"\n

\n\n

The brain must receive the full conversation, not just the current message.

Run these tests now—they should all fail:

1 pytest test_nanocode.py -v\n

\n\n

1 FAILED test_nanocode.py::test_handle_input_returns_brain_response\n2 FAILED test_nanocode.py::test_conversation_accumulates\n3 ...\n

\n\n

Good. Now let’s make them pass.

The Claude Class

Now the real brain.

The Context: We need a class that wraps the Claude API. It should handle authentication, send conversation history, and parse the response into a Thought.

The Code:

36 class Claude:\n37     \"\"\"Claude API - the brain of our agent.\"\"\"\n38 \n39     def __init__(self):\n40         self.api_key = os.getenv(\"ANTHROPIC_API_KEY\")\n41         if not self.api_key:\n42             raise ValueError(\"ANTHROPIC_API_KEY not found in .env\")\n43         self.model = \"claude-sonnet-4-6\"\n44         self.url = \"https://api.anthropic.com/v1/messages\"\n45 \n46     def think(self, conversation):\n47         headers = {\n48             \"x-api-key\": self.api_key,\n49             \"anthropic-version\": \"2023-06-01\",\n50             \"content-type\": \"application/json\"\n51         }\n52         payload = {\n53             \"model\": self.model,\n54             \"max_tokens\": 4096,\n55             \"messages\": conversation\n56         }\n57 \n58         print(\"(Claude is thinking...)\")\n59         response = requests.post(self.url, headers=headers, json=payload, timeout=120)\n60         response.raise_for_status()\n61         return self._parse_response(response.json()[\"content\"])\n

\n\n

The Walkthrough:

\nLines 39-41: Load the API key and fail fast if it’s missing.\n
\nLines 43-44: Store config. We’ll make the model configurable later.\n
\nLine 46: The think() method is the brain’s interface—same as FakeBrain.\n
\nLines 52-55: The payload includes \"messages\": conversation—the full history, not just the current message. This is the context loop.\n
\nLine 61: Parse Claude’s complex response format into our simple Thought.\n

Now the response parser:

63     def _parse_response(self, content):\n64         \"\"\"Convert Claude's response format to Thought.\"\"\"\n65         text_parts = []\n66         tool_calls = []\n67 \n68         for block in content:\n69             if block[\"type\"] == \"text\":\n70                 text_parts.append(block[\"text\"])\n71             elif block[\"type\"] == \"tool_use\":\n72                 tool_calls.append(ToolCall(\n73                     id=block[\"id\"],\n74                     name=block[\"name\"],\n75                     args=block[\"input\"]\n76                 ))\n77 \n78         return Thought(\n79             text=\"\\n\".join(text_parts) if text_parts else None,\n80             tool_calls=tool_calls\n81         )\n

\n\n

Claude’s API returns a list of “content blocks.” Each block has a type—either \"text\" or \"tool_use\". We collect all text blocks into a single string and convert tool_use blocks into ToolCall objects.

The Agent Class (Updated)

Now we update the Agent from Chapter 1 to accept a brain and maintain conversation history.

The Code:

 86 class Agent:\n 87     \"\"\"A coding agent with conversation memory.\"\"\"\n 88 \n 89     def __init__(self, brain):\n 90         self.brain = brain\n 91         self.conversation = []\n 92 \n 93     def handle_input(self, user_input):\n 94         \"\"\"Handle user input. Returns output string, raises AgentStop to quit.\"\"\"\n 95         if user_input.strip() == \"/q\":\n 96             raise AgentStop()\n 97 \n 98         if not user_input.strip():\n 99             return \"\"\n100 \n101         self.conversation.append({\"role\": \"user\", \"content\": user_input})\n102 \n103         try:\n104             thought = self.brain.think(self.conversation)\n105             text = thought.text or \"\"\n106             self.conversation.append({\"role\": \"assistant\", \"content\": text})\n107             return text\n108         except Exception as e:\n109             self.conversation.pop()  # Remove failed user message\n110             return f\"Error: {e}\"\n

\n\n

The Walkthrough:

\nLines 89-91: Accept a brain via dependency injection. Initialize an empty conversation list.\n
\nLine 101: Append the user’s message to history before calling the brain.\n
\nLines 103-107: Call the brain, extract the text, append the response to history.\n
\nLines 108-110: If the API call fails, remove the user message we just added. This keeps the conversation in a valid state.\n

Pay attention to line 101: we add the user message before calling the brain. The brain needs to see the full conversation including the current message.

\n\n

The Main Loop (Updated)

The main loop is now just a thin I/O wrapper:

115 def main():\n116     brain = Claude()\n117     agent = Agent(brain)\n118     print(\"⚡ Nanocode v0.2 (Conversation Memory)\")\n119     print(\"Type '/q' to quit.\\n\")\n120 \n121     while True:\n122         try:\n123             user_input = input(\"❯ \")\n124             output = agent.handle_input(user_input)\n125             if output:\n126                 print(f\"\\n{output}\\n\")\n127 \n128         except (AgentStop, KeyboardInterrupt):\n129             print(\"\\nExiting...\")\n130             break\n131 \n132 \n133 if __name__ == \"__main__\":\n134     main()\n

\n\n

All the logic is in the Agent class. The loop just reads input, calls handle_input(), and prints the result. This separation makes the agent testable—we test Agent.handle_input() directly without needing to mock input() or print().

Verify the Tests Pass

Run the tests again:

1 pytest test_nanocode.py -v\n

\n\n

1 test_nanocode.py::test_handle_input_returns_brain_response PASSED\n2 test_nanocode.py::test_conversation_accumulates PASSED\n3 test_nanocode.py::test_conversation_contains_correct_roles PASSED\n4 test_nanocode.py::test_brain_receives_conversation PASSED\n

\n\n

All green. The tests verify our implementation without making a single API call.

Test the Memory

Now test with the real brain:

1 python nanocode.py\n

\n\n

Try this conversation:

1 ❯ I am building a Python agent.\n2 (Claude is thinking...)\n3 \n4 That sounds exciting! Building a Python agent is a great project...\n5 \n6 ❯ What language am I using?\n7 (Claude is thinking...)\n8 \n9 You are using Python.\n

\n\n

The conversation list is doing its job.

The Context Window Problem

You might be thinking: “Can I keep this running forever?”

No.

Every loop iteration, the messages list grows:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

Turn	Approximate Tokens
1	50
10	5,000
100	50,000

Eventually, you hit the context limit—200k tokens for Claude Sonnet, 128k for DeepSeek, as low as 4k for some local models. Exceed it, and the API returns 400 Bad Request. Our error handling on line 108 catches this and reports the error, so the agent won’t crash silently. But the conversation is effectively stuck—every subsequent message will fail too, since the history is still too long.

For now, restarting the agent clears the history and gets you back on track. We’ll add proper context compaction—tracking token usage from the API response and automatically summarizing old messages before they overflow—when we build the feedback loop in Chapter 9. That’s where conversations actually blow up, and where the fix will earn its keep.

\n\n

Wrapping Up

Claude remembers now—or rather, we’ve tricked it into thinking it does. The conversation list grows with every turn, and FakeBrain lets us test the whole thing without spending a cent.

Both patterns will carry through the rest of the book. Every brain we build (Claude, DeepSeek, Ollama) will implement the same think() interface, and FakeBrain will test them all.

One loose end: our code is hardwired to Anthropic’s API. If we want to add DeepSeek or a local model, we’d have to duplicate a lot of code.

\n\n

https://martinfowler.com/articles/mocksArentStubs.html↩︎

Chapter 4: The Universal Adapter

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Adapter Pattern

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

HTTP Resilience

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Brain Interface

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The FakeBrain (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Claude Brain (Refactored)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The DeepSeek Brain

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The BRAINS Registry

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Agent Class (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Tests for Multi-Brain Support

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Main Loop (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Setting Up DeepSeek

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Try It

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

“We Just Moved Code Around”

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Part II: The Hands

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 5: The Tool Protocol

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

How Tools Actually Work

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Defining the Tool Interface

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The ReadFile Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The WriteFile Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Tool Helpers

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Updating the Thought Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Updating the Claude Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Agent Class with Tools

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Main Loop

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Test It

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 6: The Scratchpad (Memory)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The “Zero Magic” Memory

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Memory Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The ToolContext Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The SaveMemory Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Updating the Claude Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Crafting the System Prompt

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Updating the Agent Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Main Loop (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Testing Persistence

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 7: The Safety Harness (Plan Mode)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Concept

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Tests First

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Extending ToolContext

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Guarded WriteFile Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Agent Class (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Main Loop (Updated)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Testing the Harness

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Psychology of the “Plan”

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 8: The Context Pipeline (Map & Search)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The ListFiles Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The SearchCodebase Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Update the Tools List

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The “Zoom In” Test

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wait, is this RAG?

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 9: The Reality Check (Running Code)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Feedback Loop

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Tests First

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The RunCommand Tool

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Interactive Trap

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Self-Healing Demo

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The TDD Workflow

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Surgical Edit

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Closed Loop

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Context Compaction

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Security Considerations

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Part III: The Frontier

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 10: Going Dark (Local Models)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Trade-off

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Installing Ollama

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Ollama Brain Class

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Running with Ollama

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The “Infinite Loop” Experiment

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Practical Differences

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

The Hybrid Workflow

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Model Selection

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Troubleshooting Ollama

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 11: The Extension (Web Search)

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Step 1: The Meta-Prompt

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Step 2: The Surgery

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Step 3: The Reference Implementation

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Step 4: The Tests

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Self-Modification

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Wrapping Up

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n

Chapter 12: The Capstone (Building a Game)

Can Nanocode actually build something?

The “Zero Code Challenge”: build a classic Snake game using Python and Pygame. The rule—you are not allowed to write a single line of Python. You can only speak to the agent in English.

\n\n\n\n

$\"An$

Aside: This demo involves many API calls. If you hit rate limits (HTTP 429), the agent will retry automatically. For long sessions, consider using a local model via Ollama to avoid limits entirely.

Step 1: Preparation

From your project root (the directory containing nanocode.py and .env), create a working directory for the game:

1 mkdir -p snake_game\n2 cp nanocode.py snake_game/\n3 cp .env snake_game/\n4 cd snake_game\n

\n\n

If you’re using the book’s code repository, copy from ch11 instead:

1 mkdir -p snake_game\n2 cp resources/code/ch11/nanocode.py snake_game/\n3 cp .env snake_game/\n4 cd snake_game\n

\n\n

Install Pygame:

1 pip install pygame\n

\n\n\n\n\n

$\"An$

Aside: On Windows, this should work out of the box. On macOS, you may need to install SDL libraries first: brew install sdl2 sdl2_image sdl2_mixer sdl2_ttf. On Linux, install the SDL dev packages: sudo apt install libsdl2-dev libsdl2-image-dev libsdl2-mixer-dev libsdl2-ttf-dev.

Step 2: The Architect (Plan Mode)

Start the agent. We begin in plan mode because we want a blueprint before we lay bricks.

1 python nanocode.py\n

\n\n

The Prompt:

1 Build a classic Snake game using Pygame. Include a score counter and Game Over screen with a restart option. Put ALL code in ONE file: snake.py. Write the plan in PLAN.md.\n

\n\n

The agent will use write_file to create PLAN.md. Read it. It should outline the Snake class, the Food class, and the game loop—all in a single file.

If it looks good, approve it.

Step 3: The Builder (Act Mode)

Switch to act mode:

1 /mode act\n

\n\n

The Prompt:

1 Implement the plan in snake.py. All code in one file.\n

\n\n

Watch the terminal:

1   → Writing snake.py\n

\n\n

The agent is generating code based on the context it stored in PLAN.md.

Step 4: The Reality Check

Before running the game, bump the timeout. Open nanocode.py and change timeout=30 to timeout=300 in RunCommand.execute()—the default 30 seconds isn’t enough to actually play a game. (This is the one exception to the Zero Code Challenge.)

The Prompt:

1 Run the game with: python snake.py\n

\n\n

The agent executes run_command. A window pops up. You play Snake.

\n\n

If it crashes: LLMs are non-deterministic. Your agent might produce a bug on the first try. If the game crashes with an error like AttributeError: 'Snake' object has no attribute 'draw', don’t fix it yourself. Let the agent see the stderr.

The Prompt:

1 The game crashed. Read the error and fix it.\n

\n\n

The agent will read the traceback, use read_file to find the bug, use edit_file to patch it, and run it again.

\n\n

Step 5: The Pivot (Feature Creep)

The game works, but it’s ugly. The snake is just green squares. Let’s stress-test the agent’s ability to refactor.

The Prompt:

1 The game looks boring. Make the snake change color as it eats food, increase speed every 5 points, and search the web for 'cool retro game color palettes' to apply.\n

\n\n

The agent should:

\nUse search_web to find color palettes\n
\nUse read_file to understand the current rendering logic\n
\nUse edit_file to inject the new features\n
\nRun the game to verify\n

What Goes Wrong

Your results will differ from mine—LLMs are non-deterministic. But here’s what typically happens, and what to watch for.

Common failures on the first run:

\nModuleNotFoundError: No module named 'pygame' — the agent forgot you need to install it, or it ran the script in a different environment. Tell it to run pip install pygame first.\n
\nAttributeError on a method the agent defined but misspelled in the call site. The agent fixes these quickly once it sees the traceback.\n
\nOff-by-one errors in collision detection. The snake walks through walls or dies one pixel too early. These take 2-3 iterations of edit-run-fix.\n

The typical session arc:

In my tests, the agent usually gets a working (but ugly) game in 2-4 iterations. The first write produces something that crashes. The second or third fix gets it running. The feature creep step (color changes, speed ramps) adds another 3-5 iterations as the agent reads its own code, makes surgical edits, and verifies each change.

By the end, the conversation is 15-20 rounds deep. If you’re using Claude, watch the compaction trigger—around round 12-15 you’ll see “(Compacting conversation…)” as the token count approaches the 75% threshold. After compaction, the agent loses some detail about early rounds but keeps working. This is the system from Chapter 9 earning its keep.

Where the agent struggles:

Pygame’s coordinate system and event loop are tricky. The agent sometimes writes code that renders correctly but doesn’t handle keyboard input properly, or that draws the snake in the wrong order so the head appears behind the body. These are the kind of bugs a human spots instantly but the agent can’t see—it has no visual feedback, only stdout and stderr. If the game runs without errors but looks wrong, you’ll need to describe the visual bug: “The snake renders backwards—the head should be at the front.”

The point isn’t perfection on the first try. It’s that the agent converges—write, run, read the error, fix, repeat—using every tool we built across eleven chapters.

Wrapping Up

Plan, implement, crash, debug, fix, run again—that’s every chapter earning its keep.

So where does it go from here?

Epilogue

The whole thing is about 750 lines of Python. No framework. nanocode.py is yours. Do whatever you want with it:

\nGit integration that auto-commits after green tests\n
\nScreenshot-based debugging for frontend work (Claude can read images)\n
\nVoice input via Whisper so you can talk instead of type\n
\nMCP for connecting to external services your team already uses\n

Production agents like Claude Code, Cursor, and Copilot do more than this—streaming responses, parallel tool execution, tree-sitter parsing, sandboxed execution environments, multi-file context windows spanning thousands of files. The gap between 750 lines and 750,000 is real. But the architecture is the same: a brain, a loop, tools, memory, and a safety harness. Now you know what’s behind the curtain.

The models will keep getting better. The harness—the loop, the tools, the safety checks—that part is engineering. And that part isn’t going anywhere.

Acknowledgments

This content is not available in the sample book. The book can be purchased on Leanpub at https://leanpub.com/build-your-own-coding-agent.\n\n

\n\n","tocStructure",[518,519,520,521,522,523,524,525,526,527,528,529,530,531,532,533,534,535,536,537,538,539,540,541,542,543,544,545,546,547,548,549,550,551,552,553,554,555,556,557,558,559,560,561,562,563,564,565,566,567,568,569,570,571,572,573,574,575,576,577,578,579,580,581,582,583,584,585,586,587,588,589,590,591,592,593,594,595,596,597,598,599,600,601,602,603,604,605,606,607,608,609,610,611,612,613,614,615,616,617,618,619,620,621,622,623,624,625,626,627,628,629,630,631,632,633,634,635,636,637,638,639,640,641,642,643,644,645,646,647,648,649,650,651,652,653],{"_75":76,"_43":77,"_70":-5,"_78":79,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":85,"_70":-5,"_78":86,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":91,"_70":-5,"_78":92,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":94,"_70":-5,"_78":95,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":97,"_70":-5,"_78":98,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":100,"_70":-5,"_78":101,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":103,"_70":-5,"_78":104,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":107,"_70":-5,"_78":108,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":111,"_70":-5,"_78":112,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":114,"_70":-5,"_78":115,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":117,"_70":-5,"_78":118,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":120,"_70":-5,"_78":121,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":123,"_70":-5,"_78":124,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":126,"_70":-5,"_78":127,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":129,"_70":-5,"_78":130,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":132,"_70":-5,"_78":133,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":136,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":138,"_70":-5,"_78":139,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":142,"_70":-5,"_78":143,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":145,"_70":-5,"_78":146,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":148,"_70":-5,"_78":149,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":151,"_70":-5,"_78":152,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":132,"_70":-5,"_78":154,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":156,"_70":-5,"_78":157,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":159,"_70":-5,"_78":160,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":162,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":164,"_70":-5,"_78":165,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":168,"_70":-5,"_78":169,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":171,"_70":-5,"_78":172,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":174,"_70":-5,"_78":175,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":177,"_70":-5,"_78":178,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":180,"_70":-5,"_78":181,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":183,"_70":-5,"_78":184,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":186,"_70":-5,"_78":187,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":189,"_70":-5,"_78":190,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":192,"_70":-5,"_78":193,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":195,"_70":-5,"_78":196,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":198,"_70":-5,"_78":199,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":201,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":203,"_70":-5,"_78":204,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":207,"_70":-5,"_78":208,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":210,"_70":-5,"_78":211,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":213,"_70":-5,"_78":214,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":216,"_70":-5,"_78":217,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":219,"_70":-5,"_78":220,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":222,"_70":-5,"_78":223,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":225,"_70":-5,"_78":226,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":186,"_70":-5,"_78":228,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":230,"_70":-5,"_78":231,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":189,"_70":-5,"_78":233,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":235,"_70":-5,"_78":236,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":238,"_70":-5,"_78":239,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":241,"_70":-5,"_78":242,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":244,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":246,"_70":-5,"_78":247,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":250,"_70":-5,"_78":251,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":254,"_70":-5,"_78":255,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":257,"_70":-5,"_78":258,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":260,"_70":-5,"_78":261,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":263,"_70":-5,"_78":264,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":266,"_70":-5,"_78":267,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":269,"_70":-5,"_78":270,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":272,"_70":-5,"_78":273,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":275,"_70":-5,"_78":276,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":129,"_70":-5,"_78":278,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":280,"_70":-5,"_78":281,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":283,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":285,"_70":-5,"_78":286,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":289,"_70":-5,"_78":290,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":292,"_70":-5,"_78":293,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":295,"_70":-5,"_78":296,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":298,"_70":-5,"_78":299,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":272,"_70":-5,"_78":301,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":303,"_70":-5,"_78":304,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":306,"_70":-5,"_78":307,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":189,"_70":-5,"_78":309,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":311,"_70":-5,"_78":312,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":314,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":316,"_70":-5,"_78":317,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":320,"_70":-5,"_78":321,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":323,"_70":-5,"_78":324,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":326,"_70":-5,"_78":327,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":329,"_70":-5,"_78":330,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":186,"_70":-5,"_78":332,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":189,"_70":-5,"_78":334,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":336,"_70":-5,"_78":337,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":339,"_70":-5,"_78":340,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":342,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":344,"_70":-5,"_78":345,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":348,"_70":-5,"_78":349,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":351,"_70":-5,"_78":352,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":354,"_70":-5,"_78":355,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":357,"_70":-5,"_78":358,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":360,"_70":-5,"_78":361,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":363,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":365,"_70":-5,"_78":366,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":369,"_70":-5,"_78":370,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":323,"_70":-5,"_78":372,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":374,"_70":-5,"_78":375,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":377,"_70":-5,"_78":378,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":380,"_70":-5,"_78":381,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":383,"_70":-5,"_78":384,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":386,"_70":-5,"_78":387,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":389,"_70":-5,"_78":390,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":392,"_70":-5,"_78":393,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":395,"_70":-5,"_78":396,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":398,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":400,"_70":-5,"_78":401,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":404,"_70":-5,"_78":405,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":408,"_70":-5,"_78":409,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":411,"_70":-5,"_78":412,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":414,"_70":-5,"_78":415,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":417,"_70":-5,"_78":418,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":420,"_70":-5,"_78":421,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":423,"_70":-5,"_78":424,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":426,"_70":-5,"_78":427,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":429,"_70":-5,"_78":430,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":432,"_70":-5,"_78":433,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":435,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":437,"_70":-5,"_78":438,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":441,"_70":-5,"_78":442,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":444,"_70":-5,"_78":445,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":447,"_70":-5,"_78":448,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":450,"_70":-5,"_78":451,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":453,"_70":-5,"_78":454,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":456,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":458,"_70":-5,"_78":459,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":462,"_70":-5,"_78":463,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":465,"_70":-5,"_78":466,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":468,"_70":-5,"_78":469,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":471,"_70":-5,"_78":472,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":474,"_70":-5,"_78":475,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":477,"_70":-5,"_78":478,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":135,"_70":-5,"_78":480,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":84,"_43":482,"_70":-5,"_78":483,"_71":80,"_87":-5,"_88":-5,"_89":-5},{"_75":76,"_43":485,"_70":-5,"_78":486,"_71":80,"_87":-5,"_88":-5,"_89":-5},"publicChapterContent","nextChapter","sections",[],"routes/_store+/_layout","routes/_store+/_readers+/_layout","actionData","errors"]

Part II: The Hands

Table of Contents

Preface

Who This Book Is For

What You Will Build

Testing Approach

Code Examples

Conventions Used in This Book

Part I: The Brain

Chapter 1: The Zero Magic Manifesto

What is an Agent, Really?

What We Are Building

Project Setup

1. Initialize the Project

2. Create a Virtual Environment

3. Install Dependencies

4. Secure Your Keys

The AgentStop Exception

The Agent Class

Defining Success with Tests

The Main Loop

Run It

Wrapping Up

Chapter 2: The Raw Request

Get an API Key

The Vault (.env)

The Anatomy of a Request

The Headers

The Payload

The Code

Run It

Troubleshooting

Cleaning Up

Wrapping Up

Chapter 3: The Infinite Loop

The Illusion of Memory

The Testing Problem

Response Types

The FakeBrain Pattern

Defining Success

The Claude Class

The Agent Class (Updated)

The Main Loop (Updated)

Verify the Tests Pass

Test the Memory

The Context Window Problem

Wrapping Up

Chapter 4: The Universal Adapter

The Adapter Pattern

HTTP Resilience

The Brain Interface

The FakeBrain (Updated)

The Claude Brain (Refactored)

The DeepSeek Brain

The BRAINS Registry

The Agent Class (Updated)

Tests for Multi-Brain Support

The Main Loop (Updated)

Setting Up DeepSeek

Try It

“We Just Moved Code Around”

Wrapping Up

Part II: The Hands

Chapter 5: The Tool Protocol

How Tools Actually Work

Defining the Tool Interface

The ReadFile Tool

The WriteFile Tool

Tool Helpers

Updating the Thought Class

Updating the Claude Class

The Agent Class with Tools

The Main Loop

Test It

Wrapping Up

Chapter 6: The Scratchpad (Memory)

The “Zero Magic” Memory

The Memory Class

The ToolContext Class

The SaveMemory Tool