Coding Agent Harness
SaaS repo coding agent.
A coding harness track for making an AI coding agent reliable inside a real SaaS codebase: issues, tests, CI, migrations, pull requests, review, and handoff.
Define the Work Surface
Turn vague coding-agent requests into bounded code changes that can be implemented and reviewed.
Make Context Discoverable
Expose repo knowledge so the coding agent can find the right rules without flooding the prompt.
Design the Agent Interface
Treat shell commands, repo files, tools, permissions, and feedback as the coding agent's UI.
Prepare the Runway
Separate setup, environment checks, and first verification from coding work.
Limit Active Work
Prevent scope spread by forcing the coding agent to finish one behavior before starting the next.
Persist Progress
Make long-running coding work restartable without relying on chat history.
Separate Doing from Judging
Do not let the implementing coding agent be the only judge of completion.
Close the Feedback Loop
Use recurring coding-agent failures to strengthen the harness.
Instrument the Work
Make coding-agent behavior debuggable with command logs, diffs, run records, and decision artifacts.
Leave a Clean Handoff
End each coding-agent session in a state another agent or engineer can verify, understand, and continue.