How harnesses work
Three layers
- The model, such as Claude or GPT. It reads text and writes text, including requests to call tools.
- The host agent, such as Claude Code, Cursor, or Codex. It runs the model in a loop, gives it tools to read files, edit code, and run commands, and asks permission before risky actions. Some people call this layer the harness too.
- The harness, which is what this site catalogs. It is a set of files you install into the host agent that gives it a working process: when to ask questions, when to write a plan, how to test, who reviews the work, and what to write down afterwards.
A host agent without a harness follows whatever process the person prompting it follows. A harness makes that process the default and shares it across a team.
Building blocks
Harnesses are made from the same few parts. The compare page counts them for each harness.
- Skills. Markdown files, usually named
SKILL.md, each with a short description. The host loads a skill when the description matches the task, or when the user asks for it by name. Most of a harness's process lives in its skills. - Commands. Named prompts the user types, such as
/planor/review. They are explicit entry points into the process. - Subagents. Agent definitions with their own instructions and a fresh context. Harnesses use them for work that benefits from a clean slate: reviewing code, researching a codebase, or implementing one task from a plan.
- Hooks. Scripts the host runs on events, such as the start of a session or before a tool call. A harness can use a session-start hook to load its instructions into every session, so the process applies even when nobody invokes it.
- Files on disk. Plans, specs, and notes the harness writes into the repository. They are its memory: the next session, or the next person, reads them instead of starting over.
How a harness steers the agent
Most of a harness is text the model reads, and a model can ignore text. Harnesses use a few techniques to make their process stick:
- Wording. Skills written as rules ("you must write a failing test first") rather than suggestions.
- Checklists. Copying the steps of a workflow into the agent's task list, so each step is tracked and visible.
- Gates. Reviewer subagents that check the work before it counts as finished.
- Evidence. Requiring the agent to run the code and show the output before it claims success.
- Injecting hooks. Loading instructions automatically at the start of a session, instead of relying on someone to invoke them.
- Blocking hooks. Refusing a tool call, such as a destructive command, or refusing to let the turn end until a condition is met. The host runs these, so they hold even when the model ignores its instructions. Few harnesses use them; the compare page shows which.
Each entry describes which of these the harness uses, under "How it steers the agent".
Packaging and install
Most harnesses are distributed as plugins for a specific host, such as a Claude Code plugin marketplace or a Cursor plugin. Supporting another host usually means translating the files into that host's format, because skills, commands, subagents, and hooks work slightly differently in each. Some harnesses do this themselves; others rely on community ports.