agent-evaluation

Overview

Agent Skills for Context Engineering is an open collection of skills that teaches the discipline of context engineering — the art of managing a language model's context window to maximize agent effectiveness. Unlike simple prompt engineering, these skills address the holistic curation of all information entering the model's attention budget: system prompts, tool definitions, retrieved documents, message history, and tool outputs. The repository is cited in academic research from Peking University as foundational work on static skill architecture.

Key Features

• Foundational skills covering context fundamentals, degradation patterns (lost-in-middle, poisoning, distraction), and compression strategies for long-running sessions
• Architectural skills for multi-agent patterns (orchestrator, peer-to-peer, hierarchical), memory systems (short-term, long-term, graph-based), tool design, and filesystem-based context management
• Operational skills including context budgeting, agent evaluation with LLM-as-judge frameworks, planning strategies, and prompt self-optimization
• Advanced techniques for context routing, hosted background agents with sandboxed VMs, human-in-the-loop patterns, and production reliability patterns
• Platform-agnostic design — skills work across any agent platform, not tied to a specific framework

Who is this for?

Developers building AI agent systems who want to go beyond basic prompt engineering and master the science of context window management. Ideal for teams building production-grade multi-agent architectures, engineers dealing with context degradation in long-running agent sessions, and researchers studying agent skill frameworks and context engineering principles.