Try This
Run an eval against your habitat's task. Cornwall Market: take 50 real invoices, run them through the categorization eval with the agent's current rules. Score: 47/50. Sarah corrects the 3 errors — two were vendor name variations the skill did not cover, one was a new supplier. She updates the rules. Re-eval: 49/50. The remaining miss is a genuinely ambiguous item that needs human judgment.
That loop — eval, identify errors, correct, re-eval — is how the system gets better. Not through bigger models. Through better organizational knowledge captured in better skills, tested by real evals.
This is how the fourth transition works in practice. The habitat runs in production — categorizing transactions, processing invoices, surfacing anomalies. The eval system measures whether it is getting better or worse. Expert corrections feed back into the habitat's skills and memory. The eval system measures whether those corrections actually improved performance.
> "The best test environment is production. The best test harness is the tool your users actually use. Build them as the same thing."
## Check your understanding
Answer in your own words — write it down before opening the key. Your tutor grades against the criteria and generates fresh variants on retries.
**Q1.** Describe one full turn of the eval → correct → re-eval loop for a task from your own work, and state where the improvement actually comes from.