Inside the Loop

An agent's status page turns green.

The tests it ran were against a revision that was edited afterward. An export went out twice because a worker restarted between sending the request and saving the receipt. And the user's correction, "keep the public function signature unchanged," was stored in the conversation but never reached the model.

None of these failures lives in the model. A stronger model would have produced the same green status. Each one lives in the runtime: the software that builds every model input, decides which proposed actions may execute, carries them out and records what happened.

Inside the Loop is a reference on that software. Its argument is simple: the model proposes, and the runtime decides. A dependable runtime can answer five questions at any moment of any run:

• What did the model receive?
• Which operation was authorized, and by whom?
• What actually changed?
• Which evidence supports the result?
• What remains uncertain?

What's inside the second edition

• 25 chapters in four parts: Foundations, Internals, Production, and Putting It Together.
• One runtime component per chapter in Part II, each treated at equal depth: model computation, the model adapter, context construction, retrieval, memory, tools, skills, planning and control, execution state, recovery, human interaction, delegation and execution policy.
• Production chapters on the agent lifecycle, observability, evaluation, security operations, cost and capacity, scaling, and choosing a runtime language.
• 183 diagrams in one consistent visual language.
• A runnable companion experiment for Chapters 3 to 25, written in standard-library Python with no API key required. Every number the book reports as observed comes from running it.
• Source-level inspection of public implementations at pinned revisions, including LangGraph, Google ADK, Pydantic AI, smolagents, OpenClaw and Hermes, plus the author's own Rust runtime under the same scrutiny.
• Numbered invariants, a catalog of 77 named failure modes, a design-review checklist and exercises for every chapter.

Who it's for

• Engineers who build agents get mechanisms, code and the invariants a correct implementation keeps.
• Architects and reviewers get a shared vocabulary and review questions to hold a design against.
• People who decide what to build, buy or fund get each chapter's opening, which describes what goes wrong without the component in terms of cost, risk and trust.