Define context pollution in a multi-step agent and explain why it degrades reasoning quality. Describe three distinct mitigation strategies with their tradeoffs.
Free · 2 AI evals / day
Over a long run, an agent keeps adding to its **context** every single step. The user goal, every reasoning trace, every tool call, every tool result, and every error all pile up in one growing transcript. That is the agent's working memory.
Context pollution is what happens when that transcript fills with junk. Failed tool calls, stale results that are no longer true, eight identical retries, and verbose dumps all stay in the window. The useful signal gets buried under noise.
This hurts the model in two ways. First, long contexts suffer from a lost-in-the-middle effect, where the model pays most attention to the start and end and skims the middle. Second, contradictory or outdated entries can actively mislead the next step, so the agent repeats a mistake it already made.
The fixes all share one idea: stop blindly appending. You can summarise old turns into a short note, compact away dead ends, keep a clean scratchpad of facts that matter, hand subtasks to a fresh sub-agent, or retrieve only the relevant slices instead of replaying the whole history.
**Context pollution** is the gradual buildup of low-value content in a long-running agent's context window. Failed tool calls, stale observations, retry spam, and verbose dumps accumulate until noise drowns the signal.
It degrades reasoning two ways. Long contexts trigger lost-in-the-middle, where the model under-weights tokens in the middle of the window. And contradictory or outdated entries mislead the next decision, so the agent loops on a mistake it already made.
Mitigations trade fidelity against cost. Rolling summarisation compresses old turns into a short recap. It is cheap on tokens but lossy, since a summary can drop the one detail the next step needed. Context compaction drops dead ends and collapses verbose tool output to keep raw signal. A structured scratchpad or memory store holds durable facts outside the transcript, so they survive truncation. Sub-agent isolation hands a subtask to a fresh agent with a clean context and returns only the result, keeping the parent uncluttered.
The unifying principle is to retrieve over the full history rather than replay it. Treat context as a curated cache, not an append-only log.
Quick reference
TL;DR
Context pollution is the buildup of stale, noisy, or contradictory history that drowns the signal an agent needs. Fix it by compacting, summarising, and retrieving instead of appending.
Imagine taking notes on one long scroll while solving a hard problem. Every wrong turn, every dead end, every typo stays on the scroll. By hour two the scroll is enormous, and your useful notes are buried between eight crossed-out attempts and a paragraph you no longer care about. When you glance back to decide your next move, your eye lands on the failures instead of the goal, and you repeat a mistake. An agent has the same scroll: its context window. Each step appends more text, including failed tool calls and outdated results. Eventually the noise outweighs the signal, and the model reasons worse than it did at step three. The fix is to keep the scroll tidy: summarise old sections, drop dead ends, and pull back only the notes that matter right now.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Short bounded runs that finish in a handful of steps well under the window. The transcript never grows large enough to pollute, so compaction adds cost and bugs for no benefit.
Tasks where full audit fidelity of every step is a hard requirement. Summarisation and pruning destroy the verbatim trail that compliance or debugging needs.
When the bottleneck is genuinely model capability rather than accumulated noise. Trimming context cannot fix a task the model fundamentally cannot reason through.
Latency-critical single-turn calls with no accumulating history. There is no multi-step state to manage, so context curation is irrelevant overhead.
Quick scan · 2–4 min
Everything important, quickly.
Define pollution and separate it from overflow, name lost-in-the-middle and stale-entry misdirection as the two harm mechanisms, then walk summarisation, compaction, scratchpad memory, and sub-agent isolation with one tradeoff each, and close on retrieval over replay.
Real-world examples
Real products, models, and research that use this idea.
1Anthropic's Claude Code compacts long agent conversations, summarising earlier turns when the context approaches its limit so a coding session can run for hours without the goal scrolling away.
2Claude Opus 4.7 agents and Gemini 3.1 Pro both ship long context windows, yet teams still summarise and prune because raw recall degrades in the middle of very long transcripts.
3LangGraph exposes state reducers and checkpointing so a graph can trim or summarise message history per node instead of appending blindly across loop iterations.
4OpenAI's Swarm and multi-agent supervisor patterns isolate subtasks in child agents that return only a distilled result, keeping the orchestrator's context clean.
5Cursor and Cline coding agents store full file contents in a side index and inject only the relevant diff or symbol into the prompt rather than the whole repository.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow do you decide which context entries to keep versus drop without an expensive relevance model on every turn?
Sign in or go Pro to see the approach for this follow-up.
QWhat can go wrong when rolling summarisation runs on an already polluted context, and how do you guard against it?
Sign in or go Pro to see the approach for this follow-up.
QHow does sub-agent isolation actually reduce pollution rather than just moving it around?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating pollution as only a token-cost problem. The deeper harm is degraded reasoning: stale and contradictory entries actively mislead the model, not just inflate the bill.