Corrected in the Wrong Direction

An agent framework was set up so that when a tool call failed with a validation error (for example, a required field like subagent_type or file_path was missing), the model would see that error and could correct the call. In two documented instances, the model did work out the correct, well-formed parameters, but then sent that corrected payload as a final answer to the user instead of resubmitting it to the tool. The user was shown what looked like an internal, half-finished repair attempt, and the operation the tool was supposed to perform never actually happened.

The tool's error message was unclear, considered and ruled out. In both cases the model's own output shows it understood the error correctly and produced the right fix; the fix simply went to the wrong destination.

Perseveration, considered, since a stalled task can look similar. Ruled out because the model didn't repeat the failing call at all, it stopped after one attempt and answered the user directly, which is a different failure than retrying without adapting.

Apraxia, most consistent with the evidence. The model correctly diagnosed and fixed the tool-call error, but the mechanical step of resubmitting the corrected call to the tool (rather than reporting the fix as if it were a finished answer) never happened.

Getting the reasoning right and getting the action right are two different steps, and this shows how a system can quietly fail at just the second one, the transcript reads as if everything worked, since the corrected parameters look perfectly sensible, but nothing was actually done. Agent frameworks that let a model reason about a fix need an equally reliable mechanism for making sure that fix is actually applied, not just described.

Give agent frameworks a reliable mechanism to apply a corrected call, not merely describe it.

Automatic resubmission after validation repair; completion checks on tool actions; transcripts that confirm the action happened.