Can you trust an agent's own progress report while a task is still running?

Not evenly. A 2026 study tested deployed models on tau2-bench and a controlled testbed built to probe every stage of a task, and found reporting reliability depends on which stage a task has reached, not on whether the model can report at all. Two models, gpt-4.1 and gpt-5.5, score 90% or higher before acting and 87 to 89% at completion, but drop under 12% while work is underway, a mid-task blind spot between two reliable stages. The newest model tested, gpt-6-astra, wraps its stage tag inside JSON the study’s primary parser rejects and scores 0% at every stage there; only under a secondary diagnostic parser, which the authors say does not establish real compliance, does it reach 100% mid-task and 99.6% before acting, still dropping to 48.1% at completion. Both are why the study’s own authors conclude a framework should not decide whether to keep going from the model’s state report alone. A progress report that names the wrong stage is invisible the same way any other silent failure is: everything about it looks fine until it’s checked against something outside the model’s own claim.

See also: How do you know what an always-on agent did overnight?