Agent Architecture The Anatomy of an Agent Harness
Understand harness components to build robust agents.
Review your harness design against the outlined components.
Summary
This article explains that an agent is a combination of a model and a harness, where the harness provides state, tool execution, code execution, and real‑time knowledge. Core harness components include system prompts, tools, skills, MCPs, a filesystem, sandbox, orchestration logic, and middleware for deterministic execution. The harness is essential for maintaining durable state, executing code, accessing real‑time knowledge, and setting up environments. Skills allow progressive disclosure of tool descriptions, while compaction manages context rot by summarizing or offloading large tool outputs. Recursive workflows are enabled by interpreter‑driven subagent calls that keep working state outside the model context. The article also covers sandbox integration for secure code execution, memory and search for continual learning, and the importance of versioned context for rapid iteration.
Key changes
- Definition of harness as everything outside the model that enables state, code execution, and tool usage
- Core components: system prompts, tools, skills, MCPs, filesystem, sandbox, orchestration, middleware
- Skills provide progressive disclosure of tool descriptions to mitigate context rot
- Compaction summarizes or offloads large tool outputs to keep context manageable
- Interpreter‑driven recursive workflows keep working state outside the model context
- Sandbox integration offers secure, isolated code execution for agents
- Memory and search enable agents to remember and retrieve information beyond the knowledge cutoff