Agent

A model can recommend the right action. But what gives it permission to act—and what happens when the world answers back?

At the end of Reasoning, AI has produced a defensible decision supported by evidence, constraints, alternatives, and verification. Yet nothing outside the model has changed. No tool has been called, no commitment has been made, and no outcome has been observed.

Agent: How AI Takes Action begins at that boundary.

The book reveals what must exist between an intelligent conclusion and a real-world effect: a clear goal, current state, bounded authority, trustworthy tools, meaningful approval, durable execution, observation, recovery, security, and a reason to stop.

It exposes distinctions that fluent demonstrations often hide. Discovering a tool is not permission to use it. An approval is not unlimited authority. A successful API response is not proof that the intended outcome occurred. A retry can duplicate an effect. A compensation can leave consequences behind. Multiple agents can agree while sharing the same blind spot.

The book turns these risks into an understandable engineering architecture. It explores goals and success criteria, state and feedback, APIs and MCP, planning and replanning, idempotency, checkpoints, durable workflows, memory, human oversight, multi-agent coordination, prompt injection, least privilege, audit, training, evaluation, and governed execution.

Every major idea appears at three depths. Stories build intuition. Mechanisms reveal the artifacts and control boundaries. Derivations introduce the mathematics, state transitions, invariants, policies, and evaluation methods beneath them. New readers can follow the mental model, while builders and advanced practitioners can continue into implementation-level detail.

At its center is one disciplined principle:

Capability is not authority.

An agent is not merely a model with access to tools. It is a governed loop that receives an authorized goal, observes a changing environment, proposes bounded actions, crosses declared effect boundaries, verifies what happened, and knows when to continue, recover, ask, escalate, or stop.

This is Volume IV of AI, Bottom Up. Reasoning ends with a defensible decision. Agent shows how that decision becomes bounded, observable, and recoverable action.

But one successful action is not yet a dependable service.

That is where the next volume, Systems, begins.