welcome

You are buying this book at the most chaotic moment agent design has yet seen. Ask three engineers how to design an agent for the same task and you will get three starting points. One reaches for LangGraph. One opens Claude Code and starts writing skills. One reads four papers and asks ChatGPT. Each one is reasonable, and not one of them is wrong. The problem is that the field has no shared vocabulary for the design questions that decide whether an agent ships — how much context to load, what to write to memory, when to escalate to a more expensive model, when to spawn a sub-agent, what action requires human approval.

Look at how the leading harnesses are already answering those questions. Claude Code, OpenAI Codex CLI, OpenClaw, Hermes, Aider, OpenHands. Different teams, different model providers, all converging on the same architectural pieces: a small toolset, a compactable memory, a complexity router, a declarative skills layer, a permission gate. The community has started calling this whole layer the harness. The market has voted: the harness, not the model, decides whether your agent ships. What is still missing is the map of what a harness should contain.

This book is that map.

Stripped to one sentence: agent architecture is bounded resource allocation under uncertainty — the model spends, the harness budgets. The map shows where every allocation decision in your agent lives.