What causes context pollution in a long running agent?

Flashcard·Easy·4.0 · 0·~30s·Asked atAmd IBM Truera

Attempt it

Context pollution is when an agent's growing message history (tool calls, observations, intermediate thoughts) crowds out the original task or the recent relevant signal, so the model loses track or gets confused. It is caused by every step appending to the prompt without compression. Mitigations include summarising old turns, dropping irrelevant tool output, and using a separate scratchpad agent.

Your answer

An agent works by adding things to its prompt step after step. Every thought it generates, every tool call it makes, every tool result it reads, gets appended to the running message history. After 30 or 40 steps that history is huge. **Context pollution** is what happens next. The growing pile of intermediate text crowds out the original task or the most recent useful signal. The model still sees everything, but it cannot focus. The original instruction sits at the top buried under tens of thousands of words of detours and the model starts behaving as if the most recent noise is the actual goal. Two things cause it. The first is that nothing is ever removed: tool calls and results stack up forever unless someone prunes them. The second is that not all tool output is useful, so the prompt fills up with content that has no bearing on the next decision. A 50-page document the agent skimmed for one sentence still occupies the full 50 pages in context. The fixes are equally concrete. Summarise old turns into a short paragraph as the conversation gets long. Drop tool outputs that are no longer needed. Keep the original instruction near the front and pinned in a structured place. Some teams run a separate scratchpad agent that holds notes outside the main thread.

**Context pollution** is the degradation of an agent's behaviour as its accumulated message history (tool calls, results, intermediate thoughts) crowds out the original task or the most recent relevant signal. The model still receives all the tokens, but its attention is not uniform, and the more noise it has to wade through, the worse it grasps what to do next. The root cause is that every step appends to the prompt without any compression. A tool that returned a 20,000-token page of HTML stays in the context as a 20,000-token blob even when the agent only needed one fact from it. After 30 steps, the prompt is dominated by transcripts of intermediate work, and the original user instruction sits buried at the top. The failure shows up in two characteristic ways. The first is **lost-in-the-middle**: the model attends more to the start and end of the context than to the middle, so anything important that ended up in the middle stops being followed. The second is goal drift: the most recent noisy tool output starts to look more salient than the original task, and the next action chases the noise. Mitigations are well known. Summarise old turns into compact memory. Drop or truncate tool outputs that are no longer needed. Move long reference material into a retrieval store rather than the live context. Pin the original goal near the bottom of the prompt where attention is strongest.

TL;DR

Context pollution is when an agent's accumulated history of tool calls, observations, and thoughts crowds out the original task or the relevant signal, so the model loses focus or gets confused.

Memory aid

Easy to grasp

Imagine a tutor helping a student with one math problem. Every five minutes the tutor writes a fresh note: 'asked about fractions, then about decimals, then about negative numbers.' After two hours the notepad is full of side trips and the tutor can barely remember what problem they were originally solving. Context pollution is that notepad effect for an agent. Every tool call, every search result, every intermediate thought gets appended to the prompt. The original goal sits at the top buried under thousands of words of detours. When the model goes to take the next step, it spends most of its attention on the loudest noise nearby instead of the quiet original instruction.

Key concepts

Context Pollution Context Window Lost in the Middle

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Context pollution is the degradation of an agent's behaviour as its accumulated message history (intermediate thoughts, tool calls, tool results) crowds out the original task or the most recent useful signal. The model still receives every token; what changes is its ability to actually use them.

The phenomenon catches first-time agent builders because the symptoms look like a model regression. The agent worked beautifully on the first ten steps, then quietly started behaving as if it forgot the original goal. The fix is rarely a better model. It is almost always a better discipline about what to keep in the live prompt versus what to compress, drop, or move to retrieval.

This explanation builds the mental model in four steps: why a larger context window alone does not fix it, the two specific mechanisms that drive pollution, the four families of mitigation that production agents use, and a working set versus log mindset that separates a polluted agent from a clean one.

Why capacity is not the constraint

The first thing to internalise is that context pollution is not a context-window problem. Frontier models in 2026 ship with enormous windows: GPT-5.5 offers a million-token API context, Claude Opus 4.7 and Gemini 3.1 Pro support equivalent or larger working contexts. Yet long-running agents on those models still degrade after enough steps, often well before the window is anywhere near full.

The reason is that capacity and attention are different things. A model can read all of its context, but attention over that context is not uniform. The Lost in the Middle finding (Liu et al., 2023) showed that models attend most reliably to the start and end of the context, with measurably worse performance on information placed in the middle. As an agent's history accumulates, the original instruction stays at the top, the most recent step lands at the bottom, and everything in between, including older intermediate results the next step might need, falls into the lower-attention middle.

This is why buying a bigger model does not solve pollution on its own. You can fit more tokens, but the model still does not use them uniformly. Treat the window as a budget for relevant material, not as a free for all dump for everything the agent ever observed.

The two mechanisms driving pollution

Four mitigation families

Working set versus log: the right mental model

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

LangGraph offers built-in checkpointing and a message-trimming policy so long-running agents can compress old turns rather than carry them indefinitely.
Anthropic's agent guides recommend summarising prior tool results once their key fact has been extracted, to keep the active prompt focused.

What an interviewer would ask next. Try answering before peeking at the approach.

QA frontier model with a one-million-token context window still degrades on long agent runs. Why is capacity not enough?

Attention is not uniform across the window. Lost-in-the-middle shows models attend more to start and end positions than to middle. As context grows, important content gets pushed into the lower-attention middle, and salience competition between goal and recent noise grows. Capacity removes one cap; it does not remove these mechanisms.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Believing a larger context window solves it. Frontier models have a million-token context, yet attention degrades and signal to noise drops well before that limit. Pollution is a quality problem, not a capacity problem.

60 second bullets to scan on the way to the call.

Define context pollution as accumulated history crowding out the original task or recent signal.
Explain why a larger context window does not fully solve it.

Primary sources. Browse if you want the original framing.