Types of Context Rot in Harness Engineering

Description and use

<p>If you've ever worked with an LLM-powered agent long enough for a session to grow, you've probably felt it. Instructions that worked perfectly in a fresh conversation get quietly ignored 15 turns in. The agent starts referencing things it shouldn't. Retrieved snippets that were relevant three queries ago somehow bleed into an answer to a completely different question. Everything looks <em>fine</em> — no errors, no crashes — but the quality of reasoning has visibly degraded.</p> <p>That set of degradations has a name: <strong>context rot.</strong> And it's not one problem. It's five distinct

Problem

<p>If you've ever worked with an LLM-powered agent long enough for a session to grow, you've probably felt it. Instructions that worked perfectly in a fresh conversation get quietly ignored 15 turns in. The agent starts referencing things it shouldn't. Retrieved snippets that were relevant three queries ago somehow bleed into an answer to a completely different question. Everything looks <em>fine</em> — no errors, no crashes — but the quality of reasoning has visibly degraded.</p> <p>That set of degradations has a name: <strong>context rot.</strong> And it's not one problem. It's five distinct

Outcome

<p>If you've ever worked with an LLM-powered agent long enough for a session to grow, you've probably felt it. Instructions that worked perfectly in a fresh conversation get quietly ignored 15 turns in. The agent starts referencing things it shouldn't. Retrieved snippets that were relevant three queries ago somehow bleed into an answer to a completely different question. Everything looks <em>fine</em> — no errors, no crashes — but the quality of reasoning has visibly degraded.</p> <p>That set of degradations has a name: <strong>context rot.</strong> And it's not one problem. It's five distinct

Evidence

Types of Context Rot in Harness Engineering

Source