[How Coding Agents Work][15] Designing Tools an Agent Can Actually Use

How Coding Agents Work · Module 15 of 29: Designing Tools an Agent Can Actually Use Part II — The Six Layers Action layer. The highest-leverage transferable skill in the stack. Naming, schemas, descriptions, result shape, and above all the error contract — because the error message is the only recovery instruction the model ever receives. After this module you can write a tool that a model selects correctly, calls with valid arguments, and recovers from when it fails. Lessons in this module: 1. The tool interface is the whole world: name, description, schema, parameter docs 2. Single responsibility, and why action-parameter tools fail 3. Tight schemas: enums, validators, and rejecting hallucinated values 4. The error contract: fix-and-retry, retry-after, terminal 5. Result shape and size: concise defaults, pagination, truncation with a continuation hint 6. Catalogue growth: namespacing, progressive disclosure, and tool search Prerequisites: module 2. About the course: Built from the OpenAI Codex harness at commit 8d01cd42fa.