Context in. Checks out.
An agent is only as good as what you give it and what you check. I spend most of my time on those two ends, not on the prompt in the middle.
1 Context
Skills, agent roles and curated notes on the architecture and the domain, so the agent starts from how this system works, not from the internet's average.
2 Agent
One agent, one job. The agent that writes a test never writes the code that makes it pass.
3 Harness
Hooks and fixed checks around every step: a guard that sees which files each agent touched, linting, formatting, types. Guidance the agent can't skip.
4 Gates
Nothing reaches main without the full test run and a performance budget, locally and again in CI.
This site is built this way
Every change to this website went through that loop: a test list a visitor would understand, a test written by one agent, code by another, a third reviewing the whole project, and a gate of about 2.500 browser tests before each push. The tools in the Lab came out of the same way of working.
Want this in your teams?