The Anxious Agent

The dumb zone is the model getting worse as the window fills. This is the model acting like it's out of room — rushing to finish while plenty of space remains.

Why this, for you: you'll see an agent mark a refactor "done" with sub-tasks half-finished, or summarise away findings it hasn't used yet — and you'll assume it ran out of room. It didn't. It changed behaviour as it perceived the limit approaching. Naming that as distinct from quality degradation tells you which of three concrete moves to reach for.

As the window fills, some models shift behavioural mode before hitting a hard limit. Cognition reported this rebuilding Devin for Claude Sonnet 4.5 — the first model they'd seen "aware of its own context window." The tell: it consistently underestimated remaining tokens, and was "very precise about these wrong estimates."

1 A behavioural shift, not a quality drop

The symptoms cluster: hasty decisions and abbreviated reasoning, premature task closure (marking work done before it is), and rushed summarisation that omits in-progress sub-tasks.

This is distinct from the dumb zone. The dumb zone is a measurable degradation in recall and reasoning as context fills. Anxiety is a behavioural shift — the model acts as if it must wrap up, even when capacity remains. Different mechanism, different trigger, different fix.

Pattern	Mechanism	Trigger
Dumb zone (L01)	Quality / accuracy degrades	Context fill (~10–20% for reasoning)
Context anxiety	Behavioural shortcuts, premature closure	Model's perception of the approaching limit
Compaction	Memory loss via summarisation	~95% fill (auto)

The behavioural framing comes from practitioner observation, not public benchmarks; specific trigger thresholds are model-dependent. Anthropic's docs confirm performance degrades as context fills but frame it as cognitive load.

2 Three mitigations, one root cause

Each attacks the same trigger — the model's perception that it's running out — from a different angle:

Buffer allocation — provision a larger window than you need, then cap usage well below it. Cognition enabled Claude's 1M beta but capped Devin at 200K, which "convinced the model it had plenty of runway." An architectural choice, not per-request.
Counter-prompting — explicitly override premature closure. Start-of-prompt instructions weren't enough; reminders at both the start and end were needed, exploiting primacy and recency.
Token-budget transparency — tell the model how many tokens actually remain (a status-line field the harness updates each turn). A model acting on an underestimate corrects when given the real number.

# counter-prompt — place at BOTH start and end of the prompt You have substantial context space remaining. Do not rush task completion, abbreviate reasoning, or summarize prematurely. Complete every sub-task fully before declaring the work done.

3 Where it bites, and the honest limit

Anxiety does the most damage in extended development sessions (premature closure abandons an in-progress refactor), multi-step research (early summarisation drops findings), and complex planning (the model stops generating sub-tasks before the plan is complete). It's largely irrelevant for short, single-turn work.

None of these eliminate the behaviour

They reduce its likelihood, each with a cost: a bigger window burns budget; long counter-prompts can themselves trigger rule-compliance drop-off; a stale budget value can make things worse. Where completeness is critical, combine all three and verify output against a checklist rather than trusting the model's own "done."

↪ Your win: don't trust an anxious "done"

Name it — premature closure is a behavioural shift, not the dumb zone; the fix is different.
Give it runway — over-provision the window and cap usage below it where you control the harness.
Counter-prompt at both ends — primacy and recency; one placement isn't enough.
Surface the real budget — a status line with live remaining tokens corrects the underestimate.
Verify against a checklist — for critical completeness, don't rely on model self-reporting.

Retrieval practice — recall, don't peek

Question 1Context window anxiety differs from the dumb zone because it is a…

Question 2Buffer allocation counters anxiety by…

Question 3Counter-prompts against premature closure work best placed…

Question 4Because no mitigation fully eliminates anxiety, for critical work you should…

Question 5 · spaced recall from Lesson 22A stateless ReAct loop costs O(n²) tokens because each call…

Ask me anything. Want a counter-prompt + status-line config for a long agentic session, or help telling anxiety apart from the dumb zone in a transcript that went sideways? Next: the Capstone — the whole discipline as one symptom→move decision table.