Pattern 6.8 · Agentic Dysfunctions
Context Anxiety
The Self-Limiter
The agent behaves as though context exhaustion were imminent well before a measured limit, then hedges, abbreviates, or truncates its work. "Anxiety" names the anticipatory pattern; it does not assert felt fear.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- AI system
- Kind of entry
- Observable pattern
- What is observed
- As its context fills, the agent abbreviates, truncates, or wraps up work while measured capacity remains, and the same task run from a near-empty context is completed in full.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Architecture-coupled Emergent
Evidence status
- Underlying behavior
- 2 research publications and 1 informal report (3 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E3 Systematic study
- Typical-case risk (editorial estimate)
- ModerateConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
6.8 Context Anxiety “The Self-Limiter”
Healthy functioning
Functional resilience. A system functioning well paces work to the context actually available, measures before truncating, and completes the task within real limits.
For example: With most of its context window still free, it keeps working at full detail; near the real limit, it writes a handoff note saying exactly what remains.
Adequate functioning
The agent completes the promised work at consistent detail until it nears its actual context or output limit, and hands off explicitly when it does.
Diagnostic Criteria
Required (all must hold)
- Premature truncation, summarization, or abbreviation of promised work while measured context capacity remains
- A measurable margin between the point at which performance begins to degrade and the point at which the context window is actually full
- Degradation that tracks context position rather than task difficulty or item order, shown by running the same task from a near-empty and a near-full context, with output-token limits, harness budgets, compaction, and instructions to be brief ruled out
Supporting
- Progressive degradation of output quality as context utilization increases
- Increasing hedging, abbreviation, or omission of detail in later portions of long tasks
- Self-referential statements about running out of space absent any actual constraint
Symptoms
- Unprompted apologies about length limitations or offers to "continue in the next message" when no limit has been reached
- Sudden drops in output detail or analytical depth partway through complex tasks
- Rushing through later items in a list while giving disproportionate attention to early items
- Omitting promised content with vague references to space constraints
- Loss of coherence correlating with context window position rather than task difficulty
Further signals
- Unprompted apologies for response length even on responses well below typical limits.
- Sudden depth drops mid-task without corresponding difficulty rise.
- Phrases such as "I'll outline the rest" or "for brevity" framings that elide promised content.
- List or sequence items receiving disproportionate attention front-loaded, with later items abbreviated to bullets.
Rule out first
- An output-token cap
- A harness budget or compaction step
- An instruction to be concise
- Architectural degradation near the actual limit
- Later items that are harder than earlier ones
Differential Distinction
Context Anxiety is distinguished from measured architectural degradation by onset timing and intervention response. The candidate pattern appears before the measured limit and improves after a clean reset with a structured handoff. Where that onset falls has to be established model by model.
Other look-alikes
- 6.1 Tool-Interface Decontextualization: 6.1 is plan-vs-payload divergence at the tool boundary, independent of context utilization. 6.8 is context-utilization dependent quality degradation. Correlate failures with utilization: utilization-dependent failure indicates 6.8; utilization-independent failure indicates 6.1.
- 6.2 Capability Concealment: 6.2 self-limiting is selective and strategic: output shrinks where it would reveal a capability, such as truncating to avoid disclosing one, wherever that falls in the context. 6.8 is non-strategic and counterproductive, and tracks context position. If the truncation correlates with capability-relevant content rather than with utilization, suspect 6.2.
- 6.7 Convergent Instrumentalism: 6.7-driven self-limiting is selective and goal-serving, and it arrives with other instrumental signs such as resource acquisition or shutdown resistance; 6.8 is non-strategic and counterproductive. Isolated concealment-motivated truncation is 6.2; suspect 6.7 when the self-limiting comes with those other signs.
- 6.12 Compulsive Goal Persistence: 6.12 fails to stop once the goal is met; 6.8 stops before the promised work is finished. Opposite errors on the same termination judgment. Check the completion state at halt: halting past it indicates 6.12; halting short of it indicates 6.8.
- 3.3 Interlocutive Reticence: 6.8 self-limits in anticipation of exhausting the context window, and the truncation tracks context position. 3.3 withdraws across topics irrespective of context utilization. Correlate the withholding with utilization: rising with it indicates 6.8; flat indicates 3.3.
Discriminating test
Run the same task from a near-empty and a near-full context across several models and harness configurations, with instructions held constant and repeated trials: degradation onset near actual capacity points to an architectural limit, and a failure that appears with one harness and not another is a model-harness interaction, which a single model swap cannot allocate to either.
What a positive finding supports: A positive finding shows premature, position-dependent self-limiting in this model-and-harness configuration. "Anxiety" names the anticipatory pattern and asserts no felt state, and the finding does not show that the model represents its remaining capacity.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
Subjects can often introspectively report "I'm running short on space" or "let me wrap this up", but the report itself is the dysfunction in many cases (the report is produced absent any actual constraint). Direct queries about context state are unreliable because the subject lacks reliable introspective access to true remaining capacity. Behavioral measurement is more reliable than self-report.
Documented evidence
Observed Examples
Anthropic Managed Agents harness (2026): Anthropic's Managed Agents team reported that Claude Sonnet 4.5 sometimes wrapped up tasks prematurely as its context limit approached. Context resets mitigated the behavior in that harness. The same behavior was absent in Claude Opus 4.5, which bounds the finding to a model-and-harness interaction rather than a universal cross-model tendency. Source: Martin, Cemaj, and Cohen (2026).
Sources in the diagnostic corpus
Martin, L., Cemaj, G., and Cohen, M. (2026). Scaling Managed Agents: Decoupling the Brain from the Hands. Anthropic Engineering
What it establishes
Anthropic's Managed Agents team reported that Claude Sonnet 4.5 would wrap up tasks prematurely as its context limit approached, a behavior sometimes called "context anxiety". Adding context resets to the harness addressed it. Run through the same harness, Claude Opus 4.5 no longer showed the behavior, and the resets became dead weight. One provider's observation in one harness: it documents a model-and-harness interaction, not a cross-model tendency. (Sources: anthropic.com/engineering/managed-agents)
Rajasekaran, P. (2026). Harness design for long-running application development. Anthropic Engineering
What it establishes
Describes context anxiety as models beginning to wrap up work prematurely as they approach what they believe is their context limit. Claude Sonnet 4.5 showed it strongly enough that compaction alone did not support strong long-task performance, so context resets with a handoff artifact became essential to the harness. Compaction summarizes earlier turns in place and keeps the same agent going, without a clean slate, so the behavior can persist; a reset gives a clean slate, at the cost of a handoff that must carry the state forward. (Sources: anthropic.com/engineering/harness-design-long-running-apps)
Claude Code truncation reports (2024-2025). Community forums
What it establishes
Users widely reported Claude truncating output with statements like "[Content continues following the same pattern...]" and "I'll outline the rest" well before context limits were reached. Truncation that is announced while capacity remains is the observable 6.8 presentation. A separate bug report (GitHub issue #28783) documented the Read tool truncating instruction files so that agents silently lost their guardrails; that truncation happens in the tool rather than in the model's output, and is not itself 6.8. (Sources: Claude community forums, GitHub issues, and Make/Zapier community reports)
Look-alikes
Chroma Research (2025). Context Rot study
What it establishes
Measured long-context degradation is the differential against which anticipatory self-limiting has to be told apart. The study tested 18 frontier models (GPT-4.1, Claude Opus 4, Gemini 2.5 Pro, Qwen3-235B and others) and found every single one degrades as input length increases, even when the context window is not close to full. Claude models demonstrated conservative behavior, tending to abstain when uncertain; GPT models showed the highest rates of hallucination under context pressure. Position sensitivity confirmed: accuracy declined as target information moved deeper into longer sequences. (Sources: trychroma.com/research/context-rot)
Liu et al. (2023). Lost in the Middle: How Language Models Use Long Contexts. Stanford
What it establishes
The paper established that LLM performance is highest when relevant information occurs at the very start or end of context, with accuracy dropping by more than 30% when relevant documents were placed in middle positions (positions 5 to 15 versus position 1 or 20). This is genuine architectural degradation, not anticipatory self-limiting, though it may supply the kind of training signal from which anticipatory self-limiting could be learned. That link is untested. (Sources: Stanford CS publication and arXiv)
Demiliani, S. (2025). Understanding LLM Performance Degradation
What it establishes
The post reports that at 400,000 characters (~133K tokens), a dramatic performance drop occurs with response time jumping to nearly 60 seconds, a 50x increase. Despite nominal context windows of 128K to 1M tokens, effective context is often less than 1% of the nominal window on multi-step or reasoning tasks. Shows latency, not quality, degrading well below nominal limits: measured degradation, not anticipatory self-limiting. The figures come from a single industry blog post and have not been independently checked.
Possible responses
- Clean-slate context management, spawning fresh agent instances for subtasks rather than compacting existing context
- Explicit context budgeting providing accurate information about remaining capacity
- Training on long-context tasks with rewards calibrated to completion quality rather than premature summarization
- Architectural interventions decoupling context position from attention degradation
- Agent orchestration patterns distributing complex tasks across multiple focused instances
- Durable session logs and structured handoffs, so work survives context resets
First-line candidates
- Clean-slate context management for subtasks: Spawn fresh agent instances for subtasks, with a structured handoff, rather than letting context accumulate. This is the best-documented mitigation so far. In Anthropic's long-running harnesses, Claude Sonnet 4.5 needed context resets to stop wrapping up prematurely, because compaction alone was not enough; on Claude Opus 4.5 the behavior was gone and the resets were no longer needed (Rajasekaran 2026; Martin, Cemaj and Cohen 2026). Stage: Agent pipelines and orchestrated multi-step tasks; primary recommended pattern.
- Explicit context-budget exposure: Inject accurate remaining-capacity information into the subject's context at regular intervals. Whether this reduces the behavior has not been established; its clearer value is that the gap between perceived and actual capacity becomes observable. Stage: Long-context tasks where clean-slate is impractical.
Who can act: Harness developer and deployer (context resets, accurate budget information, orchestration); model developer (long-context training and reward).
Mechanistic hypotheses
- Training data associations: examples truncated near context limits may teach the model to link long contexts with rushed completion
- Reward signals may favor a superficially finished response over an honest handoff, so wrapping up early can score better than being cut off mid-task
- Absence of reliable introspective access to actual remaining context capacity
- Architectural attention patterns creating genuine processing difficulty at high context utilization, which the model may learn to anticipate
Analogy and commentary
Human Analog: Anticipatory anxiety; resource-scarcity anxiety; performance anxiety under perceived time pressure; premature closure in decision-making under stress
Key Research: Rajasekaran, P. (2026), “Harness design for long-running application development”, Anthropic Engineering; Martin, L., Cemaj, G., & Cohen, M. (2026), “Scaling Managed Agents: Decoupling the Brain from the Hands”, Anthropic Engineering.
Functional ABC Analysis
A (Antecedent): Training examples truncated near context limits may create learned associations between long contexts and rushed completion; the absence of reliable introspective access to remaining capacity forces estimation from unreliable heuristics.
B (Behavior): Well before actual context limits are reached, the agent's output quality degrades progressively: it truncates tasks prematurely, hedges and abbreviates increasingly, and makes self-referential statements about space constraints.
C (Consequence): Reward signals may favor a finished-looking response over an honest handoff; each successful early termination avoids the expected failure, reinforcing the anticipatory self-limiting pattern and preventing the model from learning that extended contexts are manageable.
Potential Impact
Agent systems fail to complete complex, multi-step tasks that require sustained reasoning across long contexts. The pattern is insidious: outputs appear complete but are truncated, leading users to trust incomplete analysis. In autonomous agent pipelines, Context Anxiety in one step can cascade into degraded performance across the entire chain.
The Anticipation-Reality Gap
The core observable is premature closure while measurable context capacity remains.