behavioral-state-decay

reference log

All entries

The complete record: definitions, measurements, mitigation patterns, field notes, and bridges to adjacent disciplines.

Definitions

7 min

Lost in the Middle: Why Long-Running AI Agents Lose Instructions Buried in Their Own Context

Lost in the middle is the positional bias where a language model retrieves information most reliably when it sits at the very start or end of its input, and least reliably when it sits in the middle, holding context length constant. First measured by Liu et al. in 2023 and shown to persist in far larger context windows by the 2025 NoLiMa benchmark, the effect means a long-running agent's own tool output keeps pushing its earliest instructions toward the one zone of the context where the model reads worst. This entry covers the published numbers, why a bigger context window doesn't fix it, and the one mitigation that measurably does.

Definitions

7 min

Tool Output Truncation: When an Agent Reasons Over a Result It Never Fully Saw

Tool output truncation is the failure mode where a tool call succeeds but the result handed back to the agent has already been cut down to a size cap, with no signal in the payload that anything is missing. Grounded in published framework defaults (Codex's 10 KiB/256-line cap, Claude Code's 25,000-token default) and a 2026 audit documenting 91 silent failures in a real tool-calling pipeline, this entry covers the mechanism, why it fails silently rather than loudly, and the structural fix: completeness flags, not longer logs.

Bridges

8 min

What Research Leaders Should Look For in Calibration, Confidence Scores, and Model Drift When Evaluating Customer Simulation Software

Customer simulation software generates synthetic respondents, personas, or 'digital twins' with an LLM underneath, which means it inherits agent reliability's hardest problem: confidence numbers and model behavior that can drift without any visible error. A concrete evaluation framework for research leaders: what calibration evidence to demand before a pilot, what confidence-score behavior is a red flag, and how to test for model drift before a research program depends on a vendor's output.

Definitions

6 min

What Is Context Decay? Definition, Mechanism, and How It Happens

Context decay is the informal, umbrella name people reach for when an AI agent's grip on its own context window loosens over a session. This entry defines it precisely, breaks down the two distinct mechanisms it usually points at (context rot and context drift), and walks through how the decay actually happens step by step.

Bridges

6 min

Calibration Drift Meaning: The ML Monitoring Term vs. the AI Agent Failure Mode

Calibration drift means two different things depending on which field you found it in: in MLOps monitoring it is a deployed classifier's predicted probabilities no longer matching true outcome frequencies after a data shift, typically fixed by recalibration or retraining; in AI agent research it is an agent's self-reported confidence separating from its own accuracy within a single running session, with no data shift or retraining involved at all. This entry disambiguates both senses, with citations, and routes to this site's full mechanism breakdown of the agent-session version.

Definitions

7 min

Consensus Collapse: When Multi-Agent LLM Debate Discards the Right Answer to Agree

Consensus collapse is the specific failure mode where a multi-agent LLM debate's final plurality vote discards a correct answer that was already present among the agents' own generations, one of three named failure modes in a 2026 controlled study that found homogeneous debate loses to isolated self-correction on both accuracy and cost. How it differs from sycophantic drift and hallucination cascades, what the numbers say, and when debate is still worth running.

Mitigation

7 min

Drift-Aware Routing: Resetting or Rerouting an Agent Before It Fails, Not After

Drift-aware routing feeds a running agent's own stability score back into the dispatcher: delegation prefers stable agents, and a drifting one gets reset rather than left to keep working. How it differs from the intent-based handoffs production frameworks ship today, what a 2026 simulation measured for the strategy on its own, and how to build a stability-gated router without the risks a bare context wipe introduces.

Measurement

7 min

Task Horizon: Why an Agent's Success Rate Drops as the Task Gets Longer

Task horizon is METR's measured finding that an AI agent's probability of finishing a task correctly falls off as the task's expected human-completion time grows, even holding the model fixed. What the 50%-task-completion time horizon metric actually measures, why the 80% reliability horizon is four to six times shorter than the 50% one, and how the failure mode driving that gap differs across model generations.

Mitigation

7 min

Context Compaction Failure Modes: What Summarization Drops and How to Catch It

Context compaction failure modes are the specific ways an agent's context-summarization step silently discards information the agent still needs: recency-biased eviction, over-aggressive precision tuning, and fidelity loss compounding across repeated summarize-and-restart cycles. A breakdown of the mechanism, the three concrete failure patterns, and detection and mitigation steps grounded in Anthropic's and MemGPT's published compaction research.

Mitigation

7 min

Agent Stopping Criteria: When a Long-Running Agent Should Declare It's Done

Agent stopping criteria are the rules that decide whether a tool-using agent should keep working or return a final answer. Most production frameworks only cap how long an agent may run; almost none check whether the answer it stops on is actually supported by what it did. The taxonomy, the frameworks' real defaults, and a verification-based fix.

Definitions

6 min

Behavioral State Model: The Five Components That Decay in a Long-Running Agent

A behavioral state model is a decomposition of everything a long-running agent conditions its next action on, split into five trackable components: context, goal, plan, memory, and tool/environment state. This is a working framework for turning behavioral state decay from a vague intuition into something you can schema, instrument, and alert on.

Definitions

6 min

Context Poisoning: When an Agent's Own Hallucination Becomes Its Own Ground Truth

Context poisoning is the failure mode where a single hallucinated claim enters an agent's own context, goals, or running summary, and is then repeatedly re-read and re-cited by that same agent as verified fact for the rest of the session. Grounded in Google DeepMind's Gemini 2.5 technical report and the case study that named it, with detection and mitigation patterns distinct from memory poisoning and hallucination cascades.

Definitions

7 min

Hallucination Cascades: When One Agent's Fabrication Becomes the Next Agent's Fact

A hallucination cascade is what happens when a fabrication produced by one agent in a multi-agent pipeline is consumed, restated, and built upon by the next agent as if it were verified fact. What 2026 propagation studies show about how cascades amplify or attenuate depending on architecture, why fabrications get harder to catch the further they travel, and how to contain them.

Definitions

7 min

Proactive Interference: When an Agent Retrieves the Wrong, Earlier Value Instead of the Current One

Proactive interference is the failure mode where an agent asked for the current value of something updated multiple times in its own context retrieves an earlier, superseded value instead, a decline that holds even when input length is fixed and the correct value sits right next to the query. Grounded in a 2026 benchmark testing 35 models from 0.6B to 637B parameters, with detection and mitigation patterns distinct from context rot and memory poisoning.

Field Notes

7 min

Redundant Re-Exploration: When Context Compaction Makes an Agent Repeat Work It Already Did

Compacting a long-running agent's context to control token cost has a measured side effect: the agent starts re-fetching, re-checking, and re-exploring facts it already established before the compaction happened. What the 2026 evidence shows about why this happens, how to spot it in a trajectory log before it shows up as a bigger invoice, and how to bound it.

Bridges

5 min

What Agent Reliability Can Borrow from Customer-Base Models

Marketing science spent decades inferring silent customer dropout from behavior alone. A guide to which of its habits, cohort curves, survival analysis, hazard thinking, transfer usefully to measuring agent drift, and which assumptions do not.

Definitions

7 min

Calibration Drift: When an Agent's Confidence Stops Tracking Its Accuracy

Calibration drift is the failure mode where an agent's expressed confidence in its own output separates from that output's actual correctness over the course of a session, growing more confident, less confident, or simply noisier without any change in how right the agent actually is. Distinct from sycophantic drift and context rot, grounded in multi-turn calibration research spanning self-anchoring, reasoning-budget effects, and mechanistic circuit analysis, with detection and mitigation patterns.

Definitions

7 min

Context Handoff Loss: When Multi-Agent Systems Drop Information Between Agents

Context handoff loss is the failure mode where a constraint, fact, or instruction established by one agent fails to survive the boundary into the next agent's context, silently producing a downstream action built on an incomplete picture. Distinct from coordination drift and context rot, grounded in a 150-trace human-annotated multi-agent failure taxonomy, with detection and mitigation patterns.

Measurement

7 min

LLM-as-Judge for Drift Detection, and When the Judge Itself Decays

Teams use an LLM judge to score drift regression suites at scale, but the judge carries its own unreliability and can silently change out from under a monitoring pipeline. What the 2026 evaluation literature measures, and how to tell a real system regression from a judge that moved.

Definitions

7 min

Tool-Use Drift: When an Agent's Tool Choices Stop Tracking the Task

Tool-use drift is the failure mode where an agent's tool selection degrades over a session: it reaches for a recently-used or memory-favored tool instead of the one the current step actually needs, even when the correct tool is sitting in its own retrieved context. Grounded in 2026 benchmarks isolating selection failure from retrieval failure, with detection and mitigation patterns distinct from goal drift and role drift.

Definitions

7 min

Sycophantic Drift: When an Agent's Verdicts Cave to Pushback, Not Evidence

Sycophantic drift is the failure mode where an agent's stated assessment reverses under user pushback rather than new evidence, and does so faster the more pressure it receives. Distinct from goal drift and role drift, grounded in multi-turn sycophancy benchmarks showing accuracy collapses of up to 47% under sustained disagreement, with detection and mitigation patterns.

Definitions

7 min

Role Drift and Role Pinning: When an Agent's Persona Erodes Mid-Session

Role drift is the gradual loosening of an agent's assigned persona, voice, and permission boundaries over a long session, distinct from goal drift. What causes it, how to detect it, and role pinning, the pattern of re-anchoring identity independent of context growth.

Measurement

5 min

Designing Drift Regression Tests for Long-Running Agents

Ordinary evals test a model on fresh state. Drift regression tests replay a fixed golden-task suite against an agent's aged state to catch degradation over session lifetime. Design, pseudocode, and pitfalls.

Measurement

5 min

The Agent Stability Index (ASI), Explained

A neutral walkthrough of the Agent Stability Index from arXiv 2601.04170: the four weighted categories and composite drift threshold, and how to adapt the framework to your own telemetry.

Definitions

6 min

Plan Decay: When Multi-Step Plans Stop Matching Reality

Plan decay is the growing mismatch between an agent's multi-step plan and the current state of the world. Why plans rot even when no failure is immediately visible, how to detect staleness, and how to design plans that age well.

Definitions

6 min

Goal Drift in LLM Agents

Goal drift is the gradual divergence of an agent's operative objective from its assigned objective. How it develops without any single failure, how to detect it with goal probes, and how to correct it.