Learn · Explained

What is an agent harness, and why does it matter?

Published 31 October 2026

Short answer. An agent harness is the software around an AI model that gives it instructions, tools, access, memory and the loop it works in. The model reasons; the harness lets it act. Claude Cowork, Microsoft Copilot agents and Claude Code are harnesses. The model decides how well an agent thinks. The harness decides what the agent can see, which systems it reaches, what it may do alone and what it can never do.

This guide explains what a harness does, names harnesses you may already use, and gives seven questions for choosing one.

What a harness does

A model on its own takes text in and gives text out. It cannot open a file, look up an invoice or remember yesterday. Anthropic describes Claude Code as “the layer around the model that provides the tools and manages the context the model sees”, and calls that layer the agentic harness. [Anthropic, How Claude Code works]

A harness does five jobs:

Job	What it means	Example
Gives the model its instructions and context	Puts the right text in front of the model at each step: the task, the instructions, the files it needs	Claude Code loads a file of project instructions, CLAUDE.md, at the start of every session
Offers tools	Lets the model read, search, write and run things, and call other systems	A tool that reads a folder, or a connection to the accounting system
Holds the access	Decides which systems and files the agent may use, and what it must ask before doing	Claude Code’s permission modes decide whether it asks before it edits a file or runs a command
Keeps memory	Carries what matters from one session to the next	Saved instructions and notes the harness loads the next time
Runs the loop	Lets the agent act, look at the result, and act again until the job is done	Run the check, read the error, correct, run it again

Anthropic describes the loop in three phases: gather context, take action, verify results. [Anthropic, How Claude Code works]

Harnesses you may already know

Each line says what the vendor’s own documentation says the product does.

Harness	Vendor	What it does, in the vendor’s words
Claude Cowork	Anthropic	Brings Claude Code’s agentic capabilities to knowledge work beyond coding: you describe an outcome and come back to finished documents, organised files or research
Microsoft Copilot agents	Microsoft	Agents that use your instructions, knowledge and actions with Copilot’s orchestrator and models, or with an orchestrator and models you bring yourself
Claude Code	Anthropic	An agentic coding tool that reads your codebase, edits files, runs commands and works with your development tools; it runs in the terminal, IDE, a desktop app and the browser
Codex	OpenAI	Inspects, edits and runs code from the terminal; you choose the model, reasoning effort, permissions and commands
Cursor	Cursor	A coding agent that works with models from OpenAI, Anthropic, Google and others
GitHub Copilot cloud agent	GitHub	Researches a repository, plans and makes code changes on a branch in its own temporary environment, for a person to review

Claude Cowork and Microsoft Copilot agents are made for office work. The other four are made for software work. Microsoft’s own description splits an agent the same way this guide does: the models are one layer, and the orchestrator that manages knowledge, tools and autonomy is another. [Microsoft Learn]

The model still matters

A stronger model reasons better, and the research is clear on that. In September 2026, researchers at EPFL and Apple gave one strong model the plainest harness they could: one session with direct access to read, write and run things. It did as well as elaborate systems with many agents and extra layers, and agents coordinating with each other did worse. The authors conclude that the model is the main driver of the results, and that elaborate layers built around a strong model give poor returns. [Brilliantov, Hernández-Cano and Abbé, 2026]

The study calls those extra layers “harnesses” too. In this guide, its plain setup is still a harness: it gave the model access, tools and a loop. Without that, the model could not have run a single line.

What the study measured was machine-learning competitions. Read If the scaffolding does not matter, what does? for the numbers and their limits.

What the harness decides

The model decides how well the agent reasons. The harness decides five other things:

What the agent sees: which instructions, records and documents reach the model at each step.
Which systems it reaches: the accounting system, the shared drive, the mailbox, or none of them.
What it may do alone: which steps the agent takes by itself, and which wait for a person.
What it can never do: the guardrails, enforced in the agent’s access and in the systems, not in the model. Read What are guardrails for an AI agent, and who enforces them?
What is recorded: whether each step, tool call and result is written to a logbook someone can read later.

An illustration: put the same model in two harnesses and ask it to propose the postings for this month’s supplier invoices. In a chat window with no access, it can explain how to post an invoice, but it has never seen yours. In a harness with read access to the invoices and the accounting system, the agent can propose each posting with the invoice as its source, for the operator to approve. The model is the same in both cases. The harness made the difference.

A harness needs your company’s knowledge

A harness out of the box knows nothing about your company. It does not know what “an open invoice” means to you, which accounts you use, or who signs off. That comes from what you put into it: your data foundation, your company dictionary, the rules for the job and the access you grant.

Read What an AI agent needs to do real work in a company for the full list.

Seven questions for choosing a harness

What can it reach? Check that it connects to the systems the job needs, through access your company grants and can withdraw. Many harnesses connect through the Model Context Protocol, an open standard for connecting AI applications to data and tools. [Model Context Protocol]
Which models does it run? Some harnesses run one provider’s models: Claude Code uses Claude models. Others offer several: Cursor works with models from OpenAI, Anthropic, Google and others. A harness that lets you change the model protects you when models change.
Which permissions does it enforce? Ask how you decide what it may do without asking. Claude Code and Codex both let you choose what the agent may do before it asks. [Anthropic; OpenAI]
Are the instructions files you own? Some harnesses keep the agent’s instructions in plain files in your own folders. Codex creates an AGENTS.md file for its instructions, and Claude Code can read the same file. Instructions in a file can move with you to another harness. [OpenAI; Anthropic]
What does it record? Ask what is logged and where. Claude Code saves each message, tool use and result to a file on the machine it runs on. GitHub Copilot cloud agent records every step in commits and logs. [Anthropic; GitHub]
Where does it run, and where does your data go? Some harnesses run on your own machine, others in the vendor’s cloud: Claude Code can run either way, and Claude Cowork runs its sessions in the cloud. The model always runs at the provider you chose, which processes what each step needs. Read the provider’s terms for what it may do with your data. [Anthropic]
Where does the operator sign off? Each step of the job needs one of the three sign-off zones: “acts alone”, “asks first”, or “the operator signs off”. Check that the harness can hold a step until the operator approves it. Read When may an AI agent act on its own?

Start from the job, then choose the harness that fits it. Your IT, in-house or your IT partner, can set up the access once you know what the job needs.

Sources

Read 31 October 2026.

Product names are trademarks of their owners, used to identify the products. No endorsement or partnership is implied.