Context Window Pollution: The Silent Agent Degradation Problem

Source: Dev.to

Introduction

Your AI agent starts sharp. Give it a task, it executes cleanly.
Give it the same task two hours later, after running continuously? It might fumble.

The Problem: Context Window Pollution

• Every tool call leaves residue in the context: partial API responses, dead‑end reasoning steps, error messages from retries, stale task context.
• The model does not cleanly distinguish “fresh instructions” from “junk from 47 steps ago.”
• Result: degraded decisions. The right answer is buried under noise.

Symptoms include:

• Agent quality degrades mid‑task (early decisions are good; later ones are sloppy).
• Unnecessary retries and re‑doing work that was already completed.
• Token usage per step grows linearly instead of staying flat.

Solution: Context Hygiene

Add this to your SOUL.md:

Every 20 steps, summarize completed work to memory/progress.md, clear intermediate reasoning, and reload only the current task context.

Benefits

• Fresh context → sharper decisions, less noise, more signal.
• Progress is preserved via the summary file—nothing important is lost, just restructured.

Recommended Context Architecture

1. Core context (always loaded)

   • Identity, rules, current task.

2. Working context (loaded per step)

   • Recent tool results, current state.

3. Archive context (on‑demand only)

   • Historical results, previous decisions.

• Only pull archive context when explicitly needed.
• Core reasoning stays clean.

Design Guidance

• Agent reliability is not just about the initial prompt; it depends on what the context looks like after many steps.
• Design for context hygiene from the beginning.

Further Resources

The Ask Patrick Library includes agent configurations with context‑flush rules built in: