Part I: One Agent
2 min read
Part I builds that unit in six stages:
- Define the agent. What an Agent Is separates the model from the harness and introduces the five design levers.
- Engineer its instructions. Structuring Agent Instructions defines the parts of a working prompt. What Prompt Research Supports separates durable findings from folklore. Managing Prompts as Software covers versioning, testing, and controlled evolution.
- Choose the model deliberately. Working With Probabilistic Models explains variance, compound error, and recurring limitations. Choosing a Model turns those constraints into a selection process.
- Design the action surface. Designing Agent Tools covers the small, unambiguous interfaces agents use best. Controlling Tool Access treats permissions as architecture. Packaging Agent Capabilities distinguishes tools, skills, and MCP servers.
- Diagnose the assembled system. Common Agent Failure Patterns shows how instructions, models, context, tools, and authority interact when an agent fails.
- Run the loop. The Agent Loop explains the inner think-act-observe cycle. Autonomous Outer Loops shows how fresh sessions and machine-verifiable outcomes turn that cycle into unattended work.
If you already run agents daily, skim this part for the failure patterns and evidence-backed design rules, then rejoin the walkthrough at Part II, where reliability engineering starts.