Designing the Feedback Loop
· Jerwin Arnado · 1 min read ·
The single biggest lever isn’t the prompt — it’s how fast and how truthfully the agent can verify its own work. Garbage feedback, garbage agent.
Verifiable surroundings
- Why ground-truth beats reasoning-in-a-vacuum (callback to part 1).
- The loop: change → run → observe → adjust.
The feedback sources, ranked
- Type checks / compile — fastest, cheapest signal.
- Tests — behavioral truth; the agent reads failures and fixes.
- Build scripts — does it actually assemble? (e.g. this site’s
bin/build.) - Screenshots / browser — the agent’s “eyes” for UI work. Frontend correctness ≠ code correctness.
Making the loop fast
- Watch mode, incremental builds, scoped test runs.
- Why a 2-second loop changes agent behavior vs a 2-minute one.
What can’t be checked, say so
- UI feel, content voice, judgment calls — the agent should flag, not fake.