Building the AI Engineering Harness a Four-Person Team Works Inside

The goal

A four-person platform team, a monorepo with several services, and a mandate to move faster with AI tools. Everyone had a coding agent within a week. The goal I set myself was to turn that from a personal productivity tool into a team practice: the agent should know how we work, be able to see what our systems are doing, and be corrected by tooling rather than by reviewers.

The approach

Encode the loops we repeat as skills the agent can run. Give it instruments, not just source code. Turn every repeated mistake into a lint rule. Keep the conventions in a file short enough that humans and agents both read it.

What I built

Skills. A ticket-lifecycle skill reads the ticket and its design notes, proposes a plan, waits for approval, and posts progress back to the tracker as work lands. A triage skill queries the log platforms across every service and environment and prints a health report. A root-cause skill takes one error from that report, traces it to source, and proposes a fix with evidence attached. A scaffolding skill generates new routes through the codebase's single request pipeline. Skills for authorization changes and for recording authenticated browser tests rounded it out.

A debugger the agent can hold. A small MCP server that manages log points. The agent injects a probe at a line, replays the failing request, reads the captured values, and removes the probe. Hypothesis, instrument, observe, revise.

Lint for agents. Four custom ESLint rules with their own tests, covering the mistakes agents make most: date handling, server-only loggers in client code, declaration style. A lint-after-edit hook runs them on every change, so the correction arrives before the pull request does.

Rules, not roles. An early experiment with a cast of named persona agents taught the opposite lesson. What shipped is a short rules file per package: conventions stated plainly, a link to the reference implementation, and the tech debt the agent must not touch.

The result

Convention comments in code review nearly disappeared, because conventions were enforced before review. Error triage went from a person's morning to a report in the channel. The rules files became the best onboarding document the team had, for humans too. Teammates picked the skills up and added their own, which is the real measure of a harness.

Takeaways

• Connect the model to logs, the tracker, and a debugger before scaling its use. Instruments beat prompts.
• The second occurrence of a mistake is a lint rule, not a sentence in a prompt.
• A harness succeeds when other people extend it.