Chapter 9The harness
The body around the brain
A harness is the software wrapped around a model. It builds the context, runs the think-act-check loop, remembers, asks your permission and connects tools, which is why the same model feels like a chatbot in one app and a coworker in another.
Harness
- Context builder
- Loop
- Memory
- Permissions
- Tool router
- Done check
On its own, a model is a brain in a jar: text in, text out. The harness gives it a body. It decides which instructions to include, which tools exist, what to remember, when to ask you before acting, and when the job is done. Anthropic calls a model with retrieval, tools and memory an "augmented LLM," and distinguishes fixed workflows from agents that decide their own next steps.1 Chat apps are light harnesses. Coding agents such as Claude Code and OpenAI's Codex are heavy harnesses that can edit files and run commands.23 Change the harness and results change: in research tests, the same model scores differently in different harnesses, and no single harness wins everywhere.456
Go deeper: kinds of harness
| Kind of harness | What it adds | Examples |
|---|---|---|
| Chat app | Instructions, memory, search, file uploads and connections to other apps. | The Claude app, with web search, Projects, memory and connectors;78910 ChatGPT.11 |
| Agentic workspace | Plans and works through many steps in a separate, sandboxed computer, asking before some risky actions. | Claude's agentic capabilities (introduced as Claude Cowork), which Anthropic's help center says run tasks in an isolated environment on its servers.1213 |
| Browser | Reads and acts on web pages. | Claude in Chrome, a browser extension that can read, click and navigate.14 |
| Coding agent | Reads a codebase, edits files, runs commands and tests. | Claude Code2 and OpenAI's Codex.3 |
| Inside other tools | An agent inside the editor you already use, with a choice of model. | GitHub Copilot offers models from several companies;15 Cursor's agent edits files and runs terminal commands.16 |
| Self-hosted | Runs on your own server; you pick the brain. | Hermes Agent, an open-source (MIT) harness from Nous Research that, the project says, works with many model providers.17 |
Notes (17)
- 1Anthropic, “Building effective agents,” December 19, 2024, https://www.anthropic.com/research/building-effective-agents (accessed October 26, 2026).
- 2Anthropic, “Claude Code overview,” Claude Code Docs, https://code.claude.com/docs/en/overview (accessed October 26, 2026).
- 3OpenAI, “openai/codex,” GitHub, https://github.com/openai/codex (accessed October 26, 2026).
- 4Jiayu Wang, Weijiang Lv, Bowen Fu, Jing Fu, Jiayi Song, Lingyu Zhang, Lanxuan Xue, et al., “Act As a Real Researcher: A Suite of Benchmarks Evaluating Frontier LLMs and Agentic Harnesses in Research Lifecycle” (preprint, arXiv, June 5, 2026), https://doi.org/10.48550/arXiv.2606.07462.
- 5Yunbei Zhang, Janet Wang, Yingqiang Ge, Weijie Xu, Jihun Hamm, and Chandan K. Reddy, “Stop Comparing LLM Agents Without Disclosing the Harness” (preprint, arXiv, May 7, 2026), https://doi.org/10.48550/arXiv.2605.23950.
- 6Yilun Yao, Xinyu Tan, Chao-Hsuan Liu, Yaoming Li, Zhengyang Wang, Zongwei Lv, Wenhan Yu, et al., “Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows” (preprint, arXiv, October 15, 2026), https://doi.org/10.48550/arXiv.2605.27922.
- 7Anthropic, “Enable and use web search,” Claude Help Center, https://support.claude.com/en/articles/10684626-enable-and-use-web-search (accessed October 26, 2026).
- 8Anthropic, “What are projects?” Claude Help Center, https://support.claude.com/en/articles/9517075-what-are-projects (accessed October 26, 2026).
- 9Anthropic, “Use Claude's chat search and memory to build on previous context,” Claude Help Center, https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context (accessed October 26, 2026).
- 10Anthropic, “Get started with custom connectors using remote MCP,” Claude Help Center, https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp (accessed October 26, 2026).
- 11OpenAI, “ChatGPT: Optimizing language models for dialogue (now 'Introducing ChatGPT'),” November 30, 2022, https://openai.com/index/chatgpt/ (accessed October 26, 2026).
- 12Anthropic, “Use Claude Cowork safely,” Claude Help Center, https://support.claude.com/en/articles/13364135-use-claude-cowork-safely (accessed October 26, 2026).
- 13Anthropic, “Claude Cowork and chat are now one Claude,” Claude blog, September 16, 2026, https://claude.com/blog/cowork-is-now-claude (accessed October 26, 2026).
- 14Anthropic, “Get started with Claude in Chrome,” Claude Help Center, https://support.claude.com/en/articles/12012173-get-started-with-claude-in-chrome (accessed October 26, 2026).
- 15GitHub, “Supported AI models in GitHub Copilot,” GitHub Docs, https://docs.github.com/en/copilot/reference/ai-models/supported-models (accessed October 26, 2026).
- 16Cursor, “Agent tools overview,” Cursor Docs, https://cursor.com/docs/agent/overview (accessed October 26, 2026).
- 17Nous Research, “hermes-agent,” Nous Research (GitHub), https://github.com/NousResearch/hermes-agent (accessed October 26, 2026).
Chapter 9 · page 1 of 2
Turn pages with the ← and → keys.