Published June 15, 2026 by gyorgy
The current discourse says you should stop prompting coding agents and start designing loops around them. Give the agent a trigger and a verifiable goal, let an evaluator check the result, and only stop when it passes. One major provider shipped a dedicated /goal command for it. People are running 25-hour unattended sessions and calling it loop engineering.
The instinct is correct. Never accept the agent's word that something is done. Demand proof.
But "verifiable" is doing all the work in that sentence. I build a framework that generates cloud infrastructure across GCP, AWS, and Azure, and I use coding agents on it daily. I keep an audit document of every serious failure. Reading it through the loop engineering lens is uncomfortable, because every bug in it would have survived a loop. Not because the loop iterated too few times. Because the verifier could not see the lie.
Here are three of them.
