Anthropic Explains Why Opus 5.5 Seems to Quit Mid-Task: Your Program Clocked It Out

A recurring complaint around Anthropic's Opus 5.5 is that the model appears to abandon long tasks halfway through, as if it decided to go home early. In a report carried by Sina's mobile news site, Anthropic's explanation is boiled down to one line: it is not the model that clocks out, it is your program that clocks out for it.

According to that explanation, the deciding factor is not the model's reasoning but the software environment the agent runs inside. When an upper-layer script, a tool call or a workflow decides the session is over, the model is cut off — and what the user experiences is a model that gave up.

That reframing matters for anyone building agents. If your evaluation only looks at final outcomes, you will count environment-driven terminations as model failures.

The result is a systematic undercount of what the model can actually do and a misleading signal for prompt and tool-chain tuning.

Anthropic's willingness to spell this out also shows how agent reliability has become a competitive front. A strong model still delivers a failed task when the harness and session management get in the way, and that gap is exactly where users form their judgment.

For teams shipping agent products, the practical move is better observability: log who or what triggered the end of a session, why it ended, and how much context was left, then count “the model stopped” separately from “the program stopped.” Only that split tells you whether to fix the prompt, the tool design or the outer workflow.

What remains unclear is how often this pattern actually occurs, and whether Anthropic will change its session and termination conventions in later releases. The signals to watch next are fuller documentation from the vendor and whether the developer community can turn “stopped halfway” from a vague complaint into a reproducible engineering bug.