M1 Field Notes · 6 min read
Design agent workflows around review capacity
Execution is becoming abundant. Good decisions and careful acceptance are still constrained.
Finished running is not finished work
A run may end successfully while producing incomplete tests, an overly broad change, or an answer that depends on an unverified assumption. Treat run completion as a handoff into review, not as an automatic done state.
Make evidence easy to scan
Ask for a short outcome summary, changed surfaces, tests and their results, screenshots when visual behavior changed, and a clear list of remaining risks. Reviewers should not need to replay an entire trace to understand the decision.
Limit work in review
Parallel execution can create a queue of unexamined output faster than a team can absorb it. Sprint and capacity planning should include review time, and concurrency limits should reflect the number of people available to make decisions.