The engineering loop
Watch an agent close the loop.
Running AI coding agents in production isn’t magic — it’s one primitive, repeated until the work is done: plan, act, observe, verify. Step through a real session and watch it happen.
A real session
One bug, one loop, stepped through
A payment test fails on CI but passes locally. Walk the agent through finding the root cause — not patching the symptom.
$ claude
> the payment test fails on CI but passes locally — find the real cause
>
plan
Plan
The agent reads the request and forms an approach before touching a single file. No fix yet — just intent.
step 1 / 7 · use ← → to step through
Four phases
The whole discipline, in one cycle
Read the request, form an approach. No edits yet — just intent and a hypothesis to test.
Gather context and make a move: read files, search the codebase, run a command, apply an edit.
Look at what came back. Did the test pass? Did the error change? Update the hypothesis.
Prove it. Run the tests. Green closes the loop; red sends it back to plan with new evidence.
Make the agent reliable at this cycle — anticipate where it goes wrong and engineer the loop so it never does again — and you’re doing harness engineering. That’s the whole game.