Context anxiety is the tendency of some models, on a lengthy task, to begin wrapping up their work prematurely as they approach what they believe is their context limit. It is named as such in Harness design for long-running application development. It appears there alongside a plainer problem with a filling context window — models losing coherence on lengthy tasks — as something only some models exhibit. What distinguishes it is that the trigger is the model’s own estimate of remaining room rather than the state of the work: the task is wrapped up early because the model believes it is running out of context, not because it is done.
Usage
The term is used in harness design to explain why one context-management technique is chosen over another. That post draws the distinction sharply against Compaction: compacting summarises earlier parts of the conversation in place so that the same agent continues on a shortened history, which preserves continuity but does not give the agent a clean slate — and so, on the post’s account, context anxiety can persist through it. A full context reset, by contrast, clears the window entirely and starts a fresh agent from a structured handoff, which the post presents as addressing both of those context-window problems at once.
The concept is therefore load-bearing in an architectural argument rather than purely descriptive: it is the stated reason a harness would accept the orchestration complexity, token overhead and latency that resets add, instead of relying on compaction alone.
When It Applies
The behaviour is reported as model-specific rather than universal, on the strength of two separate observations across different harness experiments rather than a controlled comparison. Claude Sonnet 4.5 is described as exhibiting it strongly enough that compaction alone was not sufficient for strong long-task performance, which is why context resets became essential to that harness’s design. Claude Opus 4.5 is reported to have largely removed the behaviour on its own, allowing the author to drop context resets from a later harness entirely and run agents as one continuous session with the Claude Agent SDK’s automatic compaction handling context growth.
That is also the concept’s main limitation as a design input: a harness component introduced to compensate for it can become dead weight when a newer model no longer exhibits it. The same post makes this the general case, arguing that every component in a harness encodes an assumption about what the model cannot do on its own, and that those assumptions are worth stress testing because they may be incorrect and because they can quickly go stale as models improve.
How well-established it is: the term and the observation come from one vendor’s engineering account of its own experiments, written by a member of its Labs team. No measurement of the behaviour’s frequency or magnitude is reported — the authors describe it as something they observed while decomposing why agents still tend to go off the rails on more complex tasks.
Related Terms
- Compaction — the technique this failure mode is given as persisting through
- Context rot — a different degradation associated with a filling context window
- Long-running Agent — the setting in which this behaviour matters
- Harness Engineering — the discipline in which the concept is applied