01 / Judgment
Judgment is the bottleneck.
Agents produce options quickly. People still choose the boundary, spot risk, and decide what evidence is enough.
Do thisName the decision a person owns before delegating.
02 / Evals
Define done before the work starts.
Use a test, screenshot, performance budget, or manual check. Make failure concrete before implementation.
Do thisWrite one observable failure condition first.
03 / Memory
Save decisions, not chats.
Commit repository facts. Store reusable methods separately. Record what changed and why. Drop the conversational debris.
Do thisKeep each piece of context small and scoped.
04 / Attention
More agents create more interruptions.
Parallel work helps only when the overview shows what needs a decision and hides routine activity.
Do thisLink every summary to its terminal, files, and evidence.
05 / Trust
Draw the data boundary first.
Keep data local when possible. Show sources. Capture explicitly. Make deletion clear.
Do thisDesign privacy into the flow, not the footer.
06 / Interface
Optimize for the next human decision.
Do not maximize activity. Show the diff, failed check, source, permission, or terminal that needs attention now.
The testCan a person tell what to do next?