Concepts
What is an AI harness?
A language model can write code, but it cannot run a build, read your repository, or fix the test it just broke. An AI harness is everything that closes that gap: the machinery between the model and a real computer. The model supplies judgment. The harness supplies hands, guardrails, and continuity.
Harness, agent, or IDE plugin?
The words get used interchangeably, but they name different layers. An agent is the running combination of a model plus a harness pointed at a goal — “fix this bug” in, working code out. The harness is the reusable machinery underneath: swap Claude for GPT and the same harness produces a different agent. An IDE plugin or copilot is narrower still — it suggests text into an editor, but it does not own a loop, execute tools, or carry memory of its own.
When people compare tools like Claude Code, Aider, OpenHands, or G-Rump, they are really comparing harnesses: the same frontier models are available to all of them, so the difference is entirely in the machinery.
The five components of a harness
1. The agent loop. One prompt, one answer is a chatbot. A harness runs a loop: the model acts, observes the result, and acts again until the task is done or a limit is hit. The loop needs streaming, parallelism, retries, and a ceiling — G-Rump’s runs 200 steps by default, configurable from 5 to 1,000 (architecture).
2. Tools. Tools are the verbs the model can use: read a file, run a command, query a database, take a screenshot. The breadth and quality of the tool set bounds what the agent can do at all. G-Rump ships 153 native tools (the full list), from file operations to OCR to xcodebuild.
3. Approval gates. A harness executes model-directed shell commands — the part everyone should be nervous about. Serious harnesses make the safety model explicit: per-binary exec approvals, deterministic fail-closed checks before mutating actions, protected paths the model cannot touch without sign-off (G-Rump’s security model). If a harness can’t tell you exactly what stands between the model and your machine, that is the answer.
4. Memory. Without persistence, every session starts from zero and the agent re-learns your codebase daily. A harness with memory carries context across sessions — and the interesting frontier is harnesses that learn: G-Rump records every run’s outcome, distills lessons from failures and corrections, and proposes new skills as diffs you approve (the learning loop).
5. The extension surface. No tool set is complete. A harness needs a standard way to grow — today that is MCP, the Model Context Protocol. G-Rump is both an MCP client with 67 one-click presets and an MCP server that exposes its own tools to other apps (MCP docs).
Why open source matters here
For most software, open source is a preference. For a harness it is closer to a requirement, because of what the software is: a program that runs model-directed commands on your machine, with your credentials nearby. Three things you can only truly verify with source access: what the approval gates actually check, where your API keys go, and what leaves your machine.
G-Rump’s answers are checkable: the entire app is MIT-licensed on GitHub — keys live in the macOS Keychain only, requests go straight to your provider, and there is no backend to trust. That is the standard an open-source AI harness should be held to.
Try one
The fastest way to understand a harness is to watch one work. Download G-Rump — it’s free, runs on macOS 14 or later, and takes about five minutes from download to first task (quick start). Or start from the architecture docs if you’d rather read the machinery first.