Simple Checkpoints Stop AI Assistants From Spinning in Circles Forever

Today's AI agents (software that decides and acts on its own) cram four jobs into one model: pick a plan, choose each move, write it out, and decide if the result was good enough. The problem? Nothing outside that model can say "that's wrong." So a stuck agent doesn't raise its hand — it keeps going until an outside timer or spending limit shuts it down. Picture an employee who never admits they're lost, just quietly burning the afternoon.

The fix proposed here borrows a trick from software engineering: check the work at every stage, not just at the end. Each checkpoint has two parts — a free, simple rule-based gate that needs no AI, plus an optional AI judge for fuzzy cases. Only results that pass get remembered. If something fails, you know exactly which step went wrong, and the agent stops by saying "I can't do this" instead of collapsing from exhaustion.

In a pilot test of 47 runs on ten genuinely hard research questions, the difference was stark. Agents with no verification answered none of the ten — every single run hit a step cap or ran out of allotted text. Agents with verification, but no planner, answered eight and abstained on the other two. The free rule-based gates did eight of the nine fixes, at zero added cost. Oddly, adding a planner made things worse once verification was in place.

The honest catch: this is a tiny pilot — ten questions, one small model — and it measures how agents stop, not how accurate they are at scale. The authors outline a twelve-month plan to build the full version. Still, for anyone paying for AI agents, an assistant that knows when to quit is worth real money.