Context Rot

The Story

Anthropic's own internal finding flipped the engineering instinct: for long-running tasks, resetting the context beats compressing it. A legal-tech contract-review assistant for corporate counsel showed why. The system worked section by section through long contracts, accumulating analysis as it went — party names, obligations, risk flags, unusual clauses. By the end of a 40-page contract, it had built up a comprehensive context. By the end of a 200-page master services agreement, that context was rotting in place.

The problem emerged on complex commercial agreements. By the time the assistant reached sections 25-40, its analysis quality had measurably declined. It started contradicting its own earlier findings. It confused parties -- attributing Supplier's obligations to Buyer. It missed risk flags in later sections that were identical to flags it had caught in earlier sections.

Then an associate ran an experiment. She took sections 30-40 from a problematic contract and ran them through the assistant as a fresh session -- without the accumulated analysis from sections 1-29. The quality was excellent. Same sections, same model, same prompts -- but without 28,000 tokens of accumulated prior analysis.

The assistant's context had rotted. Not because any individual piece of information was wrong, but because the accumulated volume had pushed the system past the point where additional context helped.

The Core Idea

Context rot is the progressive degradation of AI output quality as the context window fills with accumulated information over time. It's the hidden failure mode of every AI product that maintains state across multiple interactions.

Four mechanisms drive context rot. Attention dilution -- the model's attention spreads thinner as tokens increase. Information conflict -- long conversations accumulate contradictory signals. Staleness -- data loaded 30 minutes ago may be obsolete. Noise accumulation -- pleasantries, dead-end explorations, and failed tool calls compete for attention with substantive content.

Where This Hits in Production

Long-running agent workflows. An AI agent researching, drafting, then reviewing accumulates context at each step. By the review step, the window contains all prior work. The review happens with degraded attention. Anthropic's guidance: design explicit context reset points between phases rather than carrying accumulated state.

Customer support conversations. A support AI handling a complex issue over 20+ turns accumulates the full conversation history. By turn 20, attention to the original problem statement is diluted. The AI starts repeating suggestions.

Document processing pipelines. Any AI product that processes long documents section by section accumulates analytical context that eventually degrades analysis quality. Sections processed last get worse treatment than sections processed first.

In Practice: Reset vs. Compress

Here's the contract review assistant redesigned with explicit reset points -- and the results.

Remember This

1. More tokens does not mean better answers. Context rot degrades quality through four mechanisms -- attention dilution, information conflict, staleness, and noise accumulation -- and it sets in at roughly 50-60% of capacity, not at 100%.

2. Resetting beats compressing. For task-critical information, starting fresh with precisely selected context produces better outcomes than summarising accumulated context. Summaries lose the edges. The edges are where the failures live.

3. Context rot means you're paying for tokens that make your product worse. Managing context is not just quality engineering -- it is cost engineering. Frame it that way to your CFO and the investment gets approved faster.

Connecting the Dots

Context rot is to AI products what technical debt is to software. Both are invisible accumulations that degrade quality over time. Both feel manageable in the moment ("we'll clean it up later"). Both become crises when they cross a threshold. Both require proactive design to prevent -- not reactive cleanup once the damage is visible.

The eval connection. When an AI product fails an eval, the diagnostic discipline from The CONTEXT Stack points at one of the seven layers. Context rot is a different diagnosis. The layers are correctly assembled, the window correctly populated -- the failure comes from how much is in there, not what. A team that knows context rot exists asks two questions on every failure: which layer is missing or wrong? and is the context volume past the pre-rot threshold?

On the autonomy spectrum. At Levels 1-3, context rot is annoying -- the user notices and starts a new conversation. At Levels 5-7, context rot is dangerous. An autonomous agent making decisions on a rotted window might take actions informed by stale data, contradictory instructions, or attention-diluted reasoning. The agent does not know its context has rotted. It operates with the same confidence at 90% capacity as it did at 30% -- but the quality of its decisions has degraded measurably and silently.

The harness as the answer. This is why the harness exists as a discipline. The harness is the system that wraps the model and manages context lifecycle so the model never has to operate in a rotted state -- monitoring volume, triggering resets at designed thresholds, and selecting what carries forward across a phase boundary. The model stays sharp because the system around it keeps the context clean. See What is a Harness.