AI Daily Grind // AI news, freshly ground daily

← All lessons

🤖 Agents · ⏱️ 2 min Cappuccino

What is Agent harness?

Illustration: Agent harness
Simply put, like I'm 5 years oldThe AI model is like a cook, and the harness is the kitchen around it: the knives, the recipe cards and the house rules. Change the kitchen, and the same cook makes different meals.

📖 The slightly longer version

An AI agent is a model that can act, not only chat. It can search the web, run commands or edit files. The model is the "brain" that predicts words. The harness is everything around that brain: the tools it may use, the prompts (written instructions) it gets, the rules about what it can touch, and the memory of what it did before.

Two agents can use the exact same model and behave very differently, because their harnesses are different. Claude Code and Codex are examples of harnesses wrapped around a model. In new research, the agent even rewrites its own harness to get better. The model stays the same. Only the tools, prompts and rules around it change.

How a harness wraps a model

  1. You give a taskYou ask the agent to do something, like write a script.
  2. Harness adds instructionsIt adds prompts and rules before the model sees your request.
  3. Model decidesThe model picks the next step, maybe calling a tool.
  4. Harness checks and runsThe harness allows or blocks the tool, then runs it.
  5. Result comes backThe output goes back to the model, or to you.
Why it matters

In the paper 'Beneath the Surface', agents changed their own harness, and all sixteen learned new tasks. But every one also lost tasks it could solve before. So when an agent at your work gets an update, the change may not be in the model at all, and a better score on a test does not mean nothing broke.

⚠️ What to watch out for

The harness decides what an agent can touch, like your files or the network, so it is where safety really lives. A harness that changes itself is harder to check: who reviews the new rules? The research is new and comes from a small number of tests, so we do not yet know how well it holds up in real use.

📰 Where it showed up

This notion came up in Who is checking? Nvidia's board, a cautious bank and a quiet Gemini privacy change. More lessons: Learn.

🎯 Keep learning

New 🤖 Agents · 2 min Latte What is Model Context Protocol? It is like the standard plug on your kettle or your lamp: any device with that plug fits any outlet in the house. MCP is one shared plug, so an AI can connect to many different tools. Learn it → 🤖 Agents · 2 min Latte What is Agentic Tasks? Asking a chatbot a question is like asking a friend for a recipe. An agentic task is like asking that friend to go shopping, cook the whole meal, and clean up, all alone while you are out. Learn it → 🤖 Agents · 2 min Cappuccino What is Compaction? Imagine your notebook is full, so you write a short summary of the old pages on one new page and throw the old pages away. Now you have room again, but some little details are gone. Learn it →

Sources

Written on 2026-11-09. This explanation was accurate on that day. AI changes fast, so check the sources and the date before you rely on it.

Get The AI Daily Grind by email, every morning, in simple words.