Agent harness engineering for companies in Germany

What an agent harness is

An agent harness is every part of an AI agent except the model: the system prompt, the tools and MCP servers, the sandbox, the loop that runs the agent and the hooks that check its steps (LangChain, March 10, 2026). Harness engineering is the work of fitting that harness to one codebase and one team. Birgitta Böckeler of Thoughtworks sorts its controls into guides, which steer an agent before it acts, and sensors, which check its work afterwards (martinfowler.com, April 2, 2026).

Why a prompt does not bind an agent

A coding agent follows a written rule most of the time, and most of the time is too little for code that goes to production. From my client work:

• An agent reads an earlier permission as a standing one, or starts a step it had only announced.
• One harness cuts a long rule file off at a size limit and never sees the rest.
• A hook binds only the harness that loads it. An agent in a second harness passes it by.
• A check that blocks valid work gets switched off by the developers.

A rule whose violation does large damage therefore becomes a machine check, and the check gets tests of its own.

Open-source harnesses and the choice of model

DeepSeek released DeepSeek Harness under the MIT license in August 2026, as a developer preview. Its model adapter, tool registry, session log and agent loop are plugins, and the model can be a remote API or a local model server (InfoQ, August 20, 2026). An open harness lets your team choose the model, including an open-weight model on your own server. Neither an open nor a proprietary harness knows your company’s rules. That part I build.

The layers of a harness I build

The table comes from my client work. It lists what each layer stops and which harness it binds; the first two layers hold whichever harness an agent runs in.

Layer	What it stops	Binds
Limits at the resource	A write to production data, a push to the production branch	Any harness, because the database and the repository enforce them
Confirmation by a person, outside the chat	An approval quoted from the chat for a different change	Any harness
Hook before each tool call	A deletion in the file store, a push before the local CI checks have passed	The harness that loads it
Test matrix per hook	A hook that blocks valid work and gets switched off	The hook it tests
One rule file with imports	Rules that differ between Claude, Codex and Gemini	Each harness that imports it
Knowledge graph gate	A code change made without a look at the callers and earlier decisions	The harness that loads it
Review hook for longer texts	A first draft saved without review	The harness that loads it

The work page AI agents under the same rules as humans describes each layer in detail.

Who I am

I am Michael Wutzke, an AI engineer based in Frankfurt. In recent months I have learned to use AI harnesses across the full stack, from infrastructure to code and data. In client work I build software with Claude and Codex, and I am interested in open-source AI models that a company runs on its own servers. I teach at Claude Hacker House, where my sessions cover MCP servers and virtual organizations, and I hold a training authorization from IHK Frankfurt for two IT apprenticeships. Earlier I was CIO at the Frankfurt-based company Blocksize Capital. Details: Teaching and certifications.