agent context engineering

The problem

Coding agents start strong but degrade as sessions lengthen. They forget decisions, re-read files, and burn through token budgets by filling the context window with dead exploration and raw tool output.

What it does

• Establishes a three-tier memory architecture to separate standing instructions from project-long durable logs and transient session state.
• Calculates a formal token budget to prevent performance drops caused by context exhaustion.
• Implements a handoff protocol to ensure work survives session restarts and crashes without loss of state.
• Defines subagent delegation boundaries to isolate exploration noise from the main parent window.
• Identifies and prunes specific context waste like whole-file reads and redundant tool logs.

Frameworks & tools

Designed for any LLM-based coding environment including Cursor, Claude Code, and Windsurf. Focuses on context window management for high-token models like GPT-4o and Claude 3.5 Sonnet.

Why this beats prompting it yourself

Most prompts focus on what the agent should do, not what it can see. This skill provides a structural engineering approach to memory that prevents the invisible performance decay common in DIY setups, reducing long-term inference costs by identifying token waste.

Use cases

• Designing a multi-agent system where subagents must share state without polluting the parent window.
• Managing long-running refactors where decisions made in hour one must be respected in hour four.
• Recovering a failing project where the agent has become confused by its own history.
• Optimizing LLM spend by replacing whole-file reads with conclusion-based memory.

Known limitations

Does not automatically modify IDE or agent configuration files without explicit user approval. Focused on architectural strategy rather than automated script installation.