New term: agent harness

2025-09-29 · Publisher : Anthropic · Domain : Artificial Intelligence

An agent harness is everything around a model that turns it into an agent: execution loop, tools, context management, memory, permissions and checks. Anthropic popularized the term with the Claude Agent SDK in late September 2025.

What changes

  • The model alone is no longer enough: performance also depends on its harness.
  • Claude Code becomes a reusable harness packaged as a library.
  • The typical loop: gather context, act, verify, repeat.
  • The harness encompasses context engineering, tools and guardrails.
  • The word paves the way for harness engineering in 2026.

### Definition An **agent harness** is the software layer that turns a language model into an agent: the loop that plans and retries, the tools (files, terminal, web), context and memory management, permissions, the sandbox and verification. Everything that is not the model itself. ### Origin "Harness" was already used for model evaluation frameworks. Its agentic meaning spread on September 29, 2025, when Anthropic renamed the Claude Code SDK to the Claude Agent SDK. The announcement explains: "the agent harness that powers Claude Code (the Claude Code SDK) can power many other types of agents, too." On November 26, 2025, Anthropic published "Effective harnesses for long-running agents", which cemented the term. ### What the term covers - The loop described by Anthropic: gather context, take action, verify work, repeat. - Ready-made harnesses: Claude Code, Codex CLI, Gemini CLI, then personal assistants such as OpenClaw. - Recovery mechanisms for long tasks: progress files, an initializer agent, intermediate commits. ### Link with context engineering Context engineering decides what the model sees at each call. The harness encompasses it: it supplies that context, but also the tools, persistent memory, execution loop and controls that frame the agent. ### Key takeaway Two agents built on the same model can deliver very different results depending on their harness. The term shifts attention from the model to the system around it, and set the stage for harness engineering in 2026.

Links

Tags : Vocabulaire, Agents, Harness, Code