ChatGPT's agent loop got an upgrade, and it shows

ChatGPT has been quietly patching its agent loop — the mechanism that lets the model chain multiple tool calls together instead of stopping after one. The key change is an eval gate: before it commits to a next step, it asks itself whether it's confident enough to proceed. If not, it loops back for clarification rather than guessing wrong.

The idea is simple but changes the feel of the experience. Instead of plowing through a task with blind confidence and potentially wasting turns, it pauses when it's unsure — a small cost that saves bigger mistakes. The pattern echoes the bounded-confidence approach we're using elsewhere: move fast when sure, flag when not.

Why this matters for us: it's a pattern worth copying — confidence gates in tool loops save time and errors, and they don't require new models, just better wiring.

Why this matters for us: the same pattern — move when sure, flag when not — is exactly what we're building into Kelex so the agent doesn't stall or guess wrong.