Context Compaction

A technique to manage AI context overflow in long-running tasks by selectively preserving critical information while discarding less relevant details. Solves the core limitation of fixed context windows in models like claude during complex code generation.

Key Principles

• Replaces “one-shot” coding with stepwise execution (e.g., generating one function at a time instead of entire modules)
• Summarizes progress after each step (e.g., “Implemented calculate_total with test coverage”)
• Preserves decision logic while pruning redundant code/output
• Maintains task coherence across multiple agent interactions

Workflow Implementation

1. Break feature into atomic subtasks (e.g., “Write auth module → test → deploy”)
2. After each step, generate compact context summary (max 20% of original context)
3. Use summary as new context for next step
4. Never exceed context window by design

Why It Works

• Avoids context-window saturation during extended sessions
• Maintains model’s task understanding through progressive summarization
• Enables long-running agents to handle complex projects (e.g., full application builds)
• Proven effective for AI Agent code generation in production environments

Related Concepts

• AI Agent
• context-window
• ai-coding
• Stepwise Problem Solving

2026 04 14 Fixing long running Claude code sessions