Callaars & De Gruyter · AI integration & harness

The model is the easy part. The harness is not.

Anyone can call a language model. What makes an agent reliable is everything around it: what it may see, what it may do, what it may cost, and who looks over its shoulder before it does something that can't be undone. We call that the harness. We build it for you, or we help you build it yourself.

A harness around the model.

A harness is the software that surrounds an AI model and turns it into an agent you can trust. The model comes up with the next step. The harness decides whether that step is allowed.

fig. 02 · the harness around the model

Context. What information the agent gets. Not everything there is, only what it needs for this task and is allowed to see.
Tools and permissions. Which systems it may call, and with which rights. Reading the CRM is not the same as sending invoices.
Cost limit. A ceiling and an alarm point. An agent stuck in a loop otherwise costs money without anyone noticing.
Logging. Every step recorded: what it saw, what it decided, what it did. So you can show afterwards what happened.
Human check. Before anything that can't be undone, a person looks. Which steps those are is written in the spec.

Every step through the harness.

An agent works in a loop: plan, do something, check whether it's right, and carry on. The harness sits in every round of that loop. If a step falls outside what was agreed, the agent stops and reports it instead of carrying on.

fig. 03 · the loop of an agent

Three moments, one problem.

Before the build. A founder or small team that wants to build a first agent. You want to get it right the first time, not fix it later.
After the first production run. A CTO with agents already running. Something doesn't feel right. Costs are rising, or the agent did something nobody expected.
Structurally. An organisation with an AI governance team. There is a framework on paper. The technology has to follow it too.

We build it, or you do.

It always starts with a spec. After that you choose who builds.

Workshop: coding with AI. For your developers. Steering an AI coding tool within a spec, and reviewing the output for access, cost, data and accountability.
The spec. What the agent may see and do, the cost limit, where the data goes and who signs for what. Useful whatever you decide next.
We build it. The agent and the harness around it, all the way into production, with monitoring and a handover.
You build it, we check. You build within the spec. We check regularly that it still holds, and hand over the instructions for the fix.

We work every day with harnesses we built ourselves. What we build for you comes out of that practice and experience.

Already running an agent, or want one?

A 30-minute call. You tell us what's running or what you want, we tell you honestly what the harness needs to do and whether we're the right fit.

Book a 30-min call · [email protected]