Catch-up — Put the Learning Reps Back Into Agent Work
Fast agent output is useful; learning still needs deliberate reps.
- Osmani argues that agents can skip work that previously built expertise, so builders need to make those reps deliberate.
- He recommends forming a hypothesis before prompting, reading diffs, predicting failures, and sometimes working through the problem manually.
Catch-up retrospective covering sources published 28 August–31 August 2026. This is not a backdated daily edition; it pulls one durable practice from that source window.
A completed task can leave the product better while leaving your mental model untouched. That trade is fine sometimes. Repeated by default, it makes review harder precisely as agent output accelerates.
Choose one agent-assisted task this week. Before the prompt, write your expected cause and approach. After the patch, compare the diff with that prediction, name one remaining failure mode, and keep one lesson as a test or note.
Put it to work 20 minutes
A prediction-and-review note for one agent-assisted change
Run one 20-minute learning rep: predict the change, let the agent work, inspect the diff, then write what surprised you and what you would test next.
See a worked example
Use a real but bounded change. This exercise records your reasoning, not a claim that agent use has improved your skill.
- Prediction
- I expect the bug sits in state reset logic and needs one focused guard.
- Diff inspection
- The cause matched, but the patch also changed retry behavior I had not considered.
- Next test
- Test a second retry after a dropped response and verify state is not duplicated.
Your turn
Before prompting, state the likely cause and shape of the change.
Afterward, record what matched, what did not, and the evidence you saw.
Name one remaining failure mode and how you would check it.
Your draft stays in this browser. No account needed.
Do not turn every routine edit into a lesson. Slow down where the domain is new, the consequence is meaningful, or your confidence is weak.
This is attributable practitioner guidance; whether the routine improves learning for your team needs observation over repeated work.
After two weeks, can you explain and review the agent-touched area without asking the agent first?
Watch: Prediction accuracy across several tasks · Quality of issues found during human reviewKeep Going