Agents · Building
Loops you can trust
A short note on Lauren’s piece about verification, not vibes
Dennis Chong · September 2026 · ~2 min read
Getting output from agents is easy. Ask, wait, paste. The hard part is knowing whether to trust it.
Source: Lauren (@poteto) — Loops You Can Trust
Cheap output turns into slop fast: tidy diffs, confident drafts, and you still doing all the checking.
Lauren puts Andy Grove’s breakfast-factory idea on that. Toast and coffee can speed up. The plate still waits on the egg. Generation got faster. Verification did not. More agents without that check just mean more review.
I read it as architecture.
Teach the agent to check its own work against a real quality bar before you chase bigger loops. Without that, the loop makes mess. Same reason evals matter: if you cannot tell good from slop, you are hoping.
Build a verify loop you can leave alone. Reproduce. Change. Measure again. Stop when the check fails. Trust that loop, not the raw generation.
Show the work. “I fixed it” is not enough. Match evidence to what changed: failing and passing tests, before and after, a trace or screenshot.
Output got cheap. Verification is scarce. Build a verify loop with proof shaped to the change.
Loops that run are easy. Loops you can trust are worth building.