Explainer
What is an agent harness in AI?
Key takeaways
- Useful equation: Agent = Model + Harness.
- The harness is the software that keeps the model in a loop with tools, memory, sandboxes, and permissions.
- MCP is not a harness. It is a tool/context protocol many harnesses speak.
- Managed products like Claude Cowork and ChatGPT agent bundle a harness with a product UI. Hermes and OpenClaw are open harnesses you run yourself.
Table of contents
- Simple definition
- What a harness usually includes
- What a harness is not
- Why the term exploded in 2026
- FAQ
Simple definition
"Agent harness" is overloaded the same way "AI agent" is. For builders, the most useful definition is the simplest one:
An agent harness is the runtime software around a language model that turns text in/out into multi-step work: a loop, tools, memory, an execution environment, and controls.
UK AISI and much of the 2026 practitioner literature put it as agent = model + scaffolding/harness. The model reasons. The harness operates.
A simple way to think about it: the model is the brain; the harness is the body, desk, and rulebook.
What a harness usually includes
Canonical pieces (names vary by product):
- Agent loop — call the model, handle tool calls, decide when to stop or ask a human.
- Tool dispatch — file edits, browser, shell, APIs, MCP servers.
- Memory / state — what survives beyond one chat transcript.
- Sandbox / workspace — where code and files actually run.
- Context management — what gets stuffed into the window each turn.
- Guardrails — permissions, approvals, spend limits, logging.
- Channels / gateway — how humans talk to it (app UI, CLI, WhatsApp, Slack).
You do not need every subsystem on day one. A minimal harness is a loop plus a few tools. A full personal-assistant harness adds memory, messaging, and overnight runtime.
What a harness is not
- Not the model alone. Claude, GPT, Grok, or a local Llama without a loop is chat, not an agent.
- Not MCP. Model Context Protocol standardizes how tools and data connect. See Agent harness vs MCP.
- Not prompt engineering alone. Prompts sit inside the harness; they are not the runtime.
- Not every product labeled "agent." Some "agents" are thin wrappers around a single tool call.
Why the term exploded in 2026
Teams noticed two products on the same model can feel completely different. The difference is usually the harness: memory design, tool policy, sandbox, and stop conditions. Open projects like Hermes and OpenClaw made the harness itself the product. Managed vendors ship a polished harness behind Claude Cowork, ChatGPT agent, and Grok Bot.
For examples and a pick guide, see Examples of agent harnesses and Best agent harness 2026.
5 curated tools below.
ChatGPT
ChatGPT is OpenAI's flagship conversational AI, available on web, mobile, and desktop. It handles writing, coding, research, image generation, voice chat, and agentic tasks through a single interface. Paid tiers unlock frontier models, higher usage limits, Codex coding agents, Deep Research, Sora video, and team admin controls. It remains the default general-purpose AI assistant for most consumers and many businesses.
Claude Cowork
Claude Cowork is Anthropic's agentic product for non-coding knowledge work: you hand it a goal and it works across folders, connectors, and tools to deliver finished drafts for review. It is distinct from Claude Chat and from Claude Code (engineering-focused).
Grok Bot
Grok Bot is SpaceXAI's early-beta product for always-on AI teammates. Each Bot gets its own cloud computer, signs into the same apps and websites a human would use, finishes multi-step jobs end to end, and only pings you when something needs approval. Teams can run many Bots in parallel, teach workflows by demonstration, and message Bots from desktop or iOS like colleagues.
Hermes
Hermes (also called Hermes Agent) is Nous Research's open-source, self-improving AI agent you can run on desktop, terminal, or cloud. It keeps persistent memory across sessions, builds skills from completed work, and connects to messaging surfaces like Telegram, Discord, Slack, WhatsApp, Signal, and email. Distinct from the Hermes LLM model family name alone.
OpenClaw
Free, open-source autonomous AI agent that executes tasks via LLMs through messaging platforms (WhatsApp, Telegram, Discord, Slack, iMessage) with persistent memory, cron jobs, and system-level computer access.
| Tool | Best for | Pricing | Billing note |
|---|---|---|---|
| ChatGPT | AI chatbot / general-purpose assistant | Freemium | Free Trial |
| Claude Cowork | AI Agent | Paid | Paid Service |
| Grok Bot | AI Agent | Freemium | Free Trial |
| Hermes | AI Agent | Paid | Paid Service |
| OpenClaw | AI Agent | Freemium | Free Trial |
Frequently asked questions
What is an agent harness in AI?
It is the runtime around an LLM that runs the agent loop, tools, memory, sandbox, and guardrails so the model can finish multi-step work instead of only chatting.
Is Claude Code an agent harness?
Yes in the broad sense. Anthropic and others describe coding products like Claude Code as an agentic harness around a model. Personal-assistant harnesses (Hermes, OpenClaw, Cowork) emphasize persistence and channels more than an IDE session.
Do I need a harness for simple ChatGPT use?
No. Single-turn Q&A does not need one. You need a harness when the job spans tools, sessions, or unattended steps.
Agent harness vs scaffold: same thing?
Often used interchangeably. Some training literature splits "scaffold" (prompts, tools, formats the model sees) from "harness" (the execution loop). In product talk, **harness** usually means the whole non-model stack.