Context Rot

Intermediate🚫 Anti-Pattern🌀 Anti-Patterns: ContextAnthropic / Industry observation

🚫Anti-Pattern— This describes a common mistake to avoid, not a pattern to follow.

The Anti-Pattern

Agent context fills with stale, irrelevant information over time, silently degrading output quality.

Why It Happens

Agents accumulate tool outputs, past conversation turns, and intermediate results without curation. The model's attention gets diluted across thousands of tokens of noise, and it loses sight of what actually matters. The longer the session, the worse it gets — quality degrades gradually enough that nobody notices until the agent is producing garbage.

How to Fix It

Implement sliding window context management with active curation. Summarize old conversation turns instead of keeping them verbatim. Use a working memory scratchpad that the agent actively maintains. Set token budgets per context section and monitor utilization metrics. The goal is to keep the signal-to-noise ratio high throughout the entire session. The key insight is that context management isn’t optional housekeeping — it’s the single most important factor in sustained agent quality. An agent with 10K tokens of well-curated context will outperform one with 100K tokens of uncurated noise.

Diagram

  Time →
  ┌─────────────────────────────────────────┐
  │ System Prompt  │ Fresh Context │  Task  │  ← Early session (high quality)
  └─────────────────────────────────────────┘

  ┌─────────────────────────────────────────┐
  │ System │ Stale │ Stale │ Stale │ Task │  ← Late session (degraded)
  │ Prompt │ Turn  │ Tool  │ Turn  │      │
  └─────────────────────────────────────────┘

Symptoms

Agent quality degrades noticeably over long sessions
Answers become vague, contradictory, or start ignoring earlier instructions
Agent 'forgets' its own earlier solutions or decisions
Response latency increases as the conversation grows

False Positives

Short sessions that fit comfortably within the context window
Intentional context accumulation like building a knowledge base
Tasks where full conversation history is genuinely needed for accuracy

Warning Signs & Consequences

Warning Signs

Increasing latency as conversation grows — a clear telltale sign
Model referencing outdated information from earlier in the session
Decreasing relevance of responses relative to the current task
Agent contradicting its own earlier statements or decisions

Consequences

Silent quality degradation that's hard to detect without monitoring
Cascading errors as the agent builds on stale or irrelevant context
Wasted tokens processing irrelevant information in every call
User trust erodes as the agent becomes less and less useful

Remediation Steps

1Implement a sliding window that drops or summarizes old conversation turns
2Build a working memory scratchpad the agent actively curates each turn
3Set explicit token budgets per context section (system, history, tools, task)
4Add context utilization monitoring — track signal-to-noise ratio over time
5Test quality at session lengths of 10, 30, 60, and 120 minutes

Real-World Example

Coding Assistant Degradation

A coding assistant works great for the first 30 minutes — generating accurate, relevant code. But as the conversation accumulates 50+ turns of tool outputs, code snippets, and debugging back-and-forth, the agent starts generating code that contradicts its earlier solutions, forgets the project’s tech stack, and suggests libraries it already ruled out.