Glossary
what is an agent harness?
TL;DR
An agent harness is the program wrapped around an AI model that turns it into a working agent: it sends the model your request, gives it tools like reading files and running commands, runs the loop until the job is done, and keeps track of the conversation. Claude Code, Codex, Pi and Hermes are harnesses. The model is the engine; the harness is the car. Agent managers like Pane sit one level up and run several harnesses side by side.
A language model on its own can only produce text. To edit your code it needs something that reads files, runs commands, shows it the results and asks it what to do next. That something is the harness.
What a harness does
Tools, a loop, context and permissions, wrapped around a model.
Every coding agent you've used is a harness around a model. Claude Code wraps Anthropic's models, Codex wraps OpenAI's, and harnesses like Pi and Oh My Pi let you plug in dozens of providers.
| Job | What it means |
|---|---|
| Tools | Read and edit files, run shell commands, search, browse, call MCP servers. |
| The loop | Ask the model, run the tool it picked, show it the result, repeat until done. |
| Context | Decide what goes into the model's limited context window, and summarize when it fills. |
| Permissions | Ask before risky actions, or skip asking in a mode you choose. |
| Memory and resume | Save the conversation so you can pick it up later. |
How harnesses differ
Same models, very different behaviour, because the harness decides how the model works.
Two harnesses running the same model can behave very differently. They choose different tools, edit files in different formats, manage context differently and ask for permission at different moments. Oh My Pi, for example, adds a language server for safe renames and its own edit format to make weaker models edit more reliably. Pi stays deliberately small and lets you add what you need through plugins.
Some harnesses are tied to one provider, like Claude Code and Codex. Others, like Pi, Oh My Pi and Hermes, connect to many.
Harness, manager, orchestrator
The harness runs one agent. A manager runs many harnesses. An orchestrator directs them.
The words get used loosely, so here's how they stack. A harness runs one agent. An agent manager, like Pane, Orca or Conductor, runs several harnesses at once, each on its own git worktree, and shows you what they changed. An orchestrator is an agent that tells other agents what to do.
Some tools blur the lines on purpose. T3 Code calls itself an "agent harness control surface", a window over harnesses you already have. Omnigent calls itself a "meta-harness" that sits on top of other agents and adds policies and sandboxes.
Harnesses in Pane
Pane runs any harness that works in a terminal, and doesn't replace it.
Pane doesn't ship its own harness or model. It runs the harness you choose in a real terminal: Claude Code, Codex and Cursor's agent get built-in buttons, status and conversation resume, and anything else, from Gemini CLI to Pi, runs as a custom command. Your harness keeps its own login, settings and skills.
Related
Sources
- Corrections
- If anything here is wrong or out of date, tell us and we'll fix it and update the date at the top. Open an issue on GitHub
- Sources
frequently asked questions
By Parsa Khazaeepoul, co-founder of Pane. .