Smart AI Agents

Most AI agents are designed to reason first—and pay for that reasoning again every time they encounter the same problem.

Smart AI Agents presents a different architecture: use intelligence when the system encounters novelty, ambiguity, or failure, then preserve what was learned as reusable capability so future executions can take the least expensive sufficiently capable path.

The book develops this idea from architecture to working implementation. You'll learn how to separate an agent's Intelligence Plane from its Execution Plane, build executable memory, promote successful reasoning into reusable capabilities, validate deterministic execution, detect when learned capabilities become stale, and return to an LLM when the world changes.

This is not a book about eliminating LLMs. It is about eliminating unnecessary LLM usage.

The implementations are backed by reproducible experiments and public companion code. In the Chapter 3 controlled web-research benchmark, the baseline and Smart Agent architectures each completed 150 of 150 executions with 1.000 mean correctness. The baseline required 150 LLM calls. The Smart Agent required 2—avoiding LLM inference on 98.7% of executions while continuing to retrieve fresh source data.

The result was a reduction in measured model cost from $1.069980 to $0.031082 for that controlled workload, alongside substantial latency and token reductions.

Early Access: This v0.3 edition contains Chapters 1–3 and Appendix A. Future editions will progressively add the remaining implementations, experiments, benchmarks, and economic analysis as the book moves toward v1.0.

Full companion source code is available in the public GitHub repository.