Agent harness Agreed

The industry's term · its card

Definition

An agent harness turns a model into an agent, so that what the model decides is carried out in the world.

It is software that runs the loop: it gives the model its context, carries out the tool calls the model asks for, and feeds the results back until the work is done. It holds the tools, handles the context as it grows, and keeps each run it carries out. The model decides; the harness acts.

Vendor text often writes "harness" for the harness and the agent together; read it as covering both layers unless the text separates them.

Our conventions

A territory's bundle of skills, subagents, evals and hooks, such as the ki-agent-backpack repository, is not an agent harness: it runs no loop and calls no model. Under INDUSTRY-FIRST it has its own name, a backpack (HARNESS-NAME); never call it a harness.

Knowledge Islands names an agent harness by one of six values: claude-code; codex; gemini; claude-chat, Claude's own chat loop in Claude Desktop or claude.ai, Cowork included; claude-design; and chatgpt, ChatGPT's own loop, voice included. The Cairn store's roster records each agent actor's agent harness as ki-harness, beside the application it runs in as ki-host (PRESENCE-FIELDS a), so one harness keeps one value wherever it runs.

Next

The agent is a model working through an agent harness.

What each vendor calls it

Anthropic
Agentic harness (exact), Claude Code (narrower)
OpenAI
Harness (exact), Codex (narrower)
Google
Agent harness (exact)
Microsoft
Agent harness (exact), Copilot (narrower)

See the crosswalk for every term side by side.