What Is an Agent Harness? Why It Beats the Model

An agent harness is everything around an AI model that turns it into an agent that finishes work: the instructions, the context, the tools, the place the work runs, the loop that decides the next step, and the controls on what the agent may do. Microsoft's one-line version is "If the model supplies the intelligence, the harness is everything around the model" (Microsoft). It matters because of what Microsoft published on 28 October: in its own evaluations, changing models had little impact on pass rates, while changing harnesses lifted them by 10.8 points with an OpenAI model and 22.1 points with an Anthropic one (Microsoft Copilot blog).

Agent harness meaning, in plain words

The model reads and writes. On its own it cannot open a file, remember yesterday's task, retry a tool that failed, or stop before it sends something it should not. The harness does those things. An October repo that drew 132 points on Hacker News put it in a sentence: "The LLM is the brain, the harness is the body" (Hacker News).

Shrivu Shankar uses the word more broadly, for "all the infra, interfaces, context, and state that surround a stateless LLM" (The Harness Is the Company). Coding agents like Codex, Claude Code and OpenCode are harnesses in that sense, and so are personal agents. A 29 October post that drew 115 points on Hacker News called Claude Duet "the first personal-assistant-type harness that feels useful" (fig.blue, Hacker News).

What is inside an agent harness

Microsoft's Foundry team lists the parts it now runs for customers (Microsoft Foundry blog). Most harnesses have some version of each:

Instructions and skills. The job, plus reusable procedures the agent can apply across tasks.
Context. What the agent knows about this task right now, kept current as the work develops.
Tools. Search, files, code, and connections to other systems, often through MCP.
An execution environment. A sandbox or a machine where the agent can actually run things.
The agent loop. The part that coordinates model calls and tool calls toward an outcome, and decides what happens after each step.
Delegation. Handing bounded tasks to sub-agents while a main agent coordinates the result.
Controls. Authentication, approvals, limits and policy boundaries.
Recovery. Picking up after an interruption or a failed tool call instead of giving up.

The model is one line on that list. The rest is the harness.

Why everyone is talking about harnesses this week

On 28 October Microsoft renamed the GitHub Copilot harness the Microsoft Copilot harness. It now powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code, and Microsoft opened it to developers through Foundry (Microsoft). The Foundry version is in public preview "for complex, long-running tasks" (Microsoft Foundry blog). Jeff Teper and Jeff Hollan both posted it on X, Hollan with the line "Building powerful long running agents requires a powerful agent harness."

The numbers are what people are quoting. Comparing its harness with the Standard harness in Copilot Studio, Microsoft reports (Microsoft Copilot blog):

Pass rates on the same tasks rose 10.8 percentage points with an OpenAI model and 22.1 with an Anthropic model.
When a tool call failed, agents kept trying to solve the problem in 84 to 91 percent of evaluated cases, against fewer than 7 percent on the Standard harness.
The net cost of processing input fell 15 to 21 percent through better caching, even with nearly twice as many model calls.

These are Microsoft's internal evaluations and telemetry, and its footnotes say results vary by model, build and configuration. The direction is the point: same model, different harness, different result.

The idea was already moving before the announcement. "Every SaaS business will become a harness around a model" drew 168 points and 111 comments on Hacker News this month, and DeepSeek's Harness desktop app drew 417 points.

Agent harness vs model

	The model	The harness
What it does	Reasons and writes text or code	Gives the model instructions, context, tools, a place to run, and limits
What changes when you swap it	In Microsoft's tests, little	In Microsoft's tests, pass rates moved by 10.8 to 22.1 points
Who usually builds it	An AI lab	You, a vendor, or a team that runs it for you
What breaks without it	Nothing runs	The agent cannot touch files, recover from errors, or stop for approval
How often it changes	Every few months, as new models ship	Continuously, as tools, policies and work change

"Harness vs model" is not really a choice. You need both. The model sets the ceiling; the harness decides how much of that ceiling you get on real work.

Do you build your own harness?

Microsoft's own answer is that most teams should not have to. "Building one that works seamlessly across multiple models and surfaces is incredibly hard," the announcement says. "And then you have to maintain it, which is even harder" (Microsoft). The Foundry team adds that "many teams building agents end up implementing the same execution machinery," and that the work "isn't usually what differentiates your application" (Microsoft Foundry blog).

Shankar argues the opposite for the top layer. Companies, he writes, "should own the top-level harness — the one that decides what to build and reviews what comes back — and plug vendor products into it for specific workflows" (The Harness Is the Company).

Both can be true. Own the decisions and the review. Rent or hand off the machinery underneath.

A harness for knowledge work

Most people at work will never configure an agent loop. They need the harness already built around the way they work.

xSquads for Work gives each knowledge worker their own squad of specialised agents to delegate to and review, so they become 10x more productive. In harness terms, the parts are already there:

Several engines, routed by task. The app routes long reasoning, fast edits and tool calls to whichever engine fits, so you are not tied to one model.
A shared task board as the loop. The squad breaks a mission into cards that move from Backlog to To Do to In Progress to Review to Done.
Your own files as the execution environment. It runs on your computer and works on your own files and folders.
Review as the control. Finished work waits in Review until you approve it, and anything irreversible is escalated to you instead of executed.

It is on Mac now, with Windows coming soon, and free via the waitlist. If you are deciding how many agents a job needs, single agent vs multi-agent covers the trade-off, and AI agent workflow shows how a task moves through one.

A harness for shipping code

Shankar's essay also has the cleanest description of where coding harnesses are heading: "A 'software factory' is a harness whose parts are smaller harnesses — one that writes specs, one that writes code, one that reviews — plus something on top deciding what runs when" (The Harness Is the Company).

An AI software factory is a production line for code: a request comes in, agents do a bounded step, checks run, and a person decides what merges. A managed software factory is that line run for you, instead of a toolkit you assemble and staff yourself. On this site, that offer is xSquad operating as your code factory: agents open pull requests you can read, a person reviews before code is merged, nothing merges without your approval, and you add or remove agents as the workload changes. The full description is in what a managed AI software factory does, and the definition is in what is a software factory.

What to check before you trust an agent harness

Does it recover? Ask what happens when a tool call fails. Microsoft's own gap here (84 to 91 percent against under 7 percent) shows how much it varies.
Can you change the model? A harness that locks you to one model loses its value when that model falls behind.
Where does the work run? In a vendor's cloud sandbox, or on your own files and repo.
Who approves the irreversible step? Sends, payments, deletions and merges should wait for a person.
Is review built in? A harness that produces more output than anyone can read just moves the bottleneck.
What does it cost to run? Microsoft starts billing the Foundry harness runtime on 1 December 2026, on top of model inference (Microsoft Foundry blog).

For the wider risk picture, see are AI agents safe.

Where these facts come from

Read on 30 October 2026.

Introducing the Microsoft Copilot harness (Official Microsoft Blog, 28 October 2026): definition of a harness, the surfaces it powers, why building and maintaining one is hard.
The GitHub Copilot harness in Copilot Studio is now the Microsoft Copilot harness (Microsoft Copilot Blog, 28 October 2026): pass rates, resilience and cost figures.
Introducing the Microsoft Copilot harness in Foundry Agent Service (Microsoft Foundry Blog, 28 October 2026): the building blocks, the build-or-buy table, preview billing.
The Harness Is the Company (Shrivu Shankar), discussed on Hacker News.
Duet is the future of chatbot interaction (fig.blue, 29 October 2026).
Posts on X: Jeff Teper, Jeff Hollan.

Frequently Asked Questions

What is an agent harness?

An agent harness is everything around an AI model that turns it into an agent: instructions, context, tools, an execution environment, the loop that decides the next step, and the controls on what it may do. Microsoft puts it this way: if the model supplies the intelligence, the harness is everything around the model.

What is the difference between an agent harness and a model?

The model reasons and writes. The harness gives it the task, the context, the tools, a place to run and the limits it works within. In Microsoft's internal evaluations, changing models had little impact on pass rates, while changing harnesses raised them by 10.8 points with an OpenAI model and 22.1 points with an Anthropic model.

What is the Microsoft Copilot harness?

It is the harness Microsoft built at GitHub and renamed on 28 October 2026. It powers Copilot Code, Copilot Studio, the GitHub Copilot app, the CLI and VS Code, and developers can now use it in public preview through Foundry Agent Service, where Microsoft runs the agent loop and sandbox.

Do I need to build my own agent harness?

Usually not. Microsoft says building a multi-model harness is incredibly hard and maintaining it is harder, and that most teams end up rebuilding the same machinery. A common split is to own the decisions and the review, and use a harness someone else builds and maintains for the execution.

What is harness engineering?

Harness engineering is the work of building the system around the model: how the agent gets context, which tools it can call, how it recovers from failures, how it delegates to sub-agents, and where it stops for approval. It is the part that decides how much of a model's ability shows up on real work.

Is xSquads for Work an agent harness?

Yes, in the broad sense. xSquads for Work routes each piece of work to the engine that fits, runs the squad on a shared task board, works on your own files, and holds finished work in Review until you approve it. It is on Mac now, with Windows coming soon, and free via the waitlist.

What Is an Agent Harness? Why It Matters More Than the Model