Harness Engineering

You might say

The AI can edit code, but it keeps opening the wrong files and never runs tests. How do I make its working environment reliable?

Designing the rules, tools, runtime, and feedback loops that let an AI agent execute work reliably and verify the resultFor a login fix, project instructions provide the entry point, a sandbox limits writable files, and tests and CI return real results. Harness engineering is not merely choosing a stronger model or allowing more loop iterations.

Know first

AI Agent Agent Loop Tool Calling

Agent Harness EngineeringAgent Harness

Why the same agent becomes more reliable in a better working environment

Agent run · controlled by the harnessFix: login button has no loading stateSame Agent modelREADY

1Read AGENTS.md and locate the acceptance check
2Open the login component and its test
3Apply a patch inside the writable sandbox
4Run the login test and return the result

$ read AGENTS.md
Goal: add a loading state
Writable: src/auth/
Check: npm test -- login

The repository map gives the agent a real entry point, protected scope, and acceptance command.

Harness Engineering designs everything around the model that makes work executable, bounded, recoverable, and verifiable.

What the harness places around the model

Provide entry points and boundaries: Project instructions, repository maps, and permission rules tell the agent where to start, what it may edit, and which actions need approval.

Make actions executable: Files, terminals, browsers, sandboxes, and persistent state turn model proposals into controlled actions and allow interrupted work to resume.

Close the loop with feedback: Tests, CI, logs, and evaluations return real results to the next turn. Evidence—not the agent's own claim—decides whether the task is complete.

Quick checkChoose the best answer

The same coding agent often edits the wrong directory, claims a fix without testing, and forgets progress after an interruption. What should be added first?

Add a project map, controlled tools and sandbox, recoverable progress, and verification that feeds test results backMake “be careful” longer in every prompt while leaving the environment unchangedOnly raise the maximum number of loops so the agent can keep guessing

You can say this to an AI Agent

Read the project instructions first and list the files related to login. Make changes only inside the auth sandbox. After every change, run the login test and use failures as input to the next turn. When it passes, report the changed files, command, and result instead of only saying “done.”

Learn next

1Agent Loop→2Human-in-the-loop→3CI→4End-to-End Test→