Field guides for checking agent work.

01Purpose

Practical guides for reviewing and verifying work from coding agents.

Coding agents now open pull requests, run tools and report their own results. The useful questions are operational. What separates an agent from an assistant? Which permissions should it get? How do you know a passing check belongs to the exact code under review? How should cost and speed be measured without counting only the successes?

Each guide answers one of those questions with a method a reviewer can apply, and states where that method stops.

02Method

Worked examples, with their limits stated.

Examples keep illustrative numbers and authored fixtures apart from measured results, automated checks apart from human acceptance, and a successful run apart from a general claim. Where a guide uses an invented example, it says so.

TraceAxiom builds Verify, a GitHub Action in development that runs a pull request's tests on base and head, checks its claims and posts one verdict. The guides do not depend on it, and none of them reports a Verify result.

• Illustrative examples are labelled
• Limitations appear beside conclusions
• No invented run becomes a product result
• Automated checks never stand in for human acceptance

03Reading path

Start with the agent, then follow the evidence.

Begin with AI coding agents for the execution model, then the security guide for permissions around files, networks, credentials and spending. The verification guide covers how to check an AI-generated pull request against the exact code it changes.

For review work, follow the guides on CI artifacts, runtime evidence, sandbox cleanup and MCP tool permissions. The evaluation, cost and speed guides show how to compare workflows while counting failed attempts, verification and review.