What is an agent harness?

An agent harness is the system around a language model that decides when an agent runs, what it receives and where its work goes next. The model produces one step, and the harness supplies the schedules, events, skills, tickets and reviews that turn steps into finished work.

How it works

A harness answers three plain questions for every agent. What wakes it? What does it get to read? Where does its output go?

The agents I work with wake in different ways. Some run on a schedule and check for tickets in the morning. Some react to an event: a ticket, a comment, a build. Some run in loops that look for useful work, and some work toward a goal for 24 or 36 hours.

The work then moves through stages that the harness fixes in advance. For example, the transcript of a meeting lands in a folder of a repository and a monitoring process notices it. An agent reads it with a skill written for that kind of file and extracts decisions and tasks. Those become tagged tickets. Other agents react to the tags: one updates documentation, another implements, another reviews the result.

None of those steps needs me at the keyboard. The laptop is only an interface.

Where it shows up in real work

My own agents already run for long periods without me, and I noticed that meetings were the place where I was still the integration layer. I would close a call, write notes, copy context, create a ticket and point an agent at the right repository.

So I am building what I call a physical agent harness: a small portable station with an iPad mini, an external camera, a microphone, a light, a USB-C hub and a battery. The conversation is transcribed, and an agent can work out which client and workspace it belongs to, which decisions were made and what must become a ticket. The meeting stops being text that I process later. It becomes an event that enters the same infrastructure as the rest of the work.

The physical harness captures what agents cannot normally see: human conversation, intent and decisions. The software harness handles what happens next.

What it is not

A harness is not the model, and it is not a single agent. It is also not a promise of full autonomy. Some checks stay hard to automate, such as judging whether an interface feels wrong, and the answer to those is usually one more layer in the harness, with a person who reviews the important output.

When you need one

You need a harness when agents already do useful work but a person still carries every result from one place to the next. Each manual boundary that repeats is a candidate for a system. Look for the moments where you explain the same context twice, or where work waits until you return to the computer.

Work and evidence

The Agent Graph Runtime (opens in a new tab)

My contribution: Authored the runtime model separating graph construction, qualification, admission, execution and evidence.

What this supports: A published conceptual model for reasoning about executable identity and workflow lifecycle.

What it does not establish: A model and reference representation, not a throughput or reliability guarantee.

Source reviewed:

Common questions

What is a physical agent harness?

It is the hardware boundary between human activity and a system of agents. Mine is a portable meeting station with a camera, a microphone, a light and a battery, so every conversation is captured the same way wherever I am. Its job is to hand the conversation to the software as an event, so the work continues after I hang up.

What wakes an agent inside a harness?

A schedule, an event or a loop. Some agents wake at a set hour and check for tickets, some react to a comment or a build, and some run continuously looking for useful work. Others work toward a longer goal for a day or more.

Does an agent harness make the work fully autonomous?

No. The aim is that every repeated manual step becomes a candidate for a system, not that everything runs alone overnight. My own role moves upward: I define intent, review the important output, make decisions and talk to people.

Are static workflows and dynamic planners different kinds of harness?

My preprint The Agent Graph Runtime argues that those labels mix up four separate questions: when the graph is built, where it becomes fixed and addressable, who authors it and what runs it. Answered separately, static, dynamic and hybrid systems turn out to be points on one map. The paper proposes a model and excludes any claim that it is better or more general than others.

Open to senior AI engineering roles and contract engagements, remote in the US, Mexico and Latin America.