Execution Feedback

The difference between an agent and a plan is that an agent gets to see what happened. A tool returns a result, an error, or nothing at all, and the next decision is made with that in hand. Systems that act without reading their own output are open-loop, and they fail in the way open-loop systems always have.

Feedback quality decides agent quality more than model quality does. A stack trace is a usable signal, a generic failure string is not, and truncating a long output at the wrong place removes the part carrying the diagnosis. Code as the action format wins partly for this reason — the interpreter's errors are precise, structured, and already designed to be read.

Each cycle conditions the next action on the accumulated transcript, so an agent's behaviour is a policy over histories rather than states. That means an uninformative observation does not merely waste a step — it leaves the belief unchanged, and repeating the same action from the same belief is the loop these systems fall into.