For CTOs and engineering leaders
Harness engineering:
what makes agents reliable
An AI model on its own only writes text. The harness around it turns it into an agent that works reliably in your codebase: rules, tools, permissions and checks. We build it for your team and work with one ourselves every day.
- Rules CLAUDE.md, AGENTS.md
- Tools Skills, MCP servers
- Permissions Set centrally
- Checks Tests, a second agent
What is harness engineering?
An AI agent consists of the AI model, such as Claude or GPT, and its harness: the software and files that steer the model, give it tools and check its results. Harness engineering means building this harness for a team and improving it over time. In 2026, Birgitta Böckeler of Thoughtworks put it this way: agent equals model plus harness.
- The AI model
- Claude, GPT or Gemini. It proposes the next step, nothing more.
- The harness
- The loop that calls the model again and again, plus rules, tools, permissions and checks. Claude Code and Codex come with the basics built in.
- Harness engineering
- Building on those basics for a team: with your rules, your systems and your checks.
Prompt, context or harness: what is the difference?
| Term | What it is about | What it includes |
|---|---|---|
| Prompt engineering | A single request to the model | A well-written task, examples, a target format |
| Context engineering | What is in the model's context at each step | Instructions, tools, data and the history so far, selected step by step |
| Harness engineering | The lasting environment in which an agent works, session after session | Lasting files, permissions and checks that grow with the team |
Context engineering picks the context for the next step, harness engineering builds the environment for every session.
What a harness is made of and what ours looks like.
How our pages are built.
The harness is not a concept on slides. This is how we build the pages of this website, including the one you are reading.
Research
Several agents research in parallel: the term, the searches, the competition and our own proof. Another checks the key facts against primary sources.
Brief and copy
The research turns into a brief and the copy. The language and brand rules are ready as skills and apply the same way to every page.
Build
One agent builds the page from existing components and checks it on mobile and desktop.
Review
Fresh agents with no prior context read the copy and the code, round after round, until no serious issue is left.
Approval
Ralf reads the page himself in the browser. Only then does it go live through a pull request.
From prompt to harness.
- July 2025
Context engineering
LangChain describes how to select an agent's context step by step.
- November 2025
Harnesses for long tasks
Anthropic shows how an agent works for many hours: with a progress log, git history and tests in the browser.
- Early 2026
The OpenAI report
A team builds an internal product without writing code by hand, with about 1,500 pull requests in five months. People steer, agents execute.
- April 2026
Agent equals model plus harness
Birgitta Böckeler of Thoughtworks gives the term its formula and separates steering before from checking after.
Three terms and how they relate.
The craft
Agentic engineering
Building software with AI agents. Our offer: we place engineers who work exactly this way.
The environment
Harness engineering
The part of the craft that builds and maintains the environment the agents work in.
This pageThe role
Forward Deployed Engineer
Who does this at the client: a senior engineer inside the client's team who writes production code and sets up the harness.
Who builds your harness?
A harness takes shape in the repository, while working on real tasks. We offer two ways to get there.
Roll out Claude Code to your team
We set up rules, skills, permissions and checks, and your team keeps working with them.
Add an engineer to your repository
The engineers we place build the harness in your repository while they work on your tasks.
Common questions.
Is harness engineering the same as agentic engineering or a forward deployed engineer?
No. The three are related but mean different things: the craft, the environment and the role. Harness engineering matters to you as soon as several people on your team work with agents and the results have to be reliable.
Is a CLAUDE.md or AGENTS.md enough?
It is the start. A rules file steers the agent before it acts. Only the checks afterwards show whether the result is right: tests, type checks, a second agent. A harness needs both.
Do we need our own harness engineer?
Usually not as a permanent role. Someone who has done it before sets up the harness, and then your team maintains it day to day. We roll it out with Claude Code or place an engineer who builds it in your repository.
How is a harness different from an automation?
An automation runs fixed steps. An agent decides for itself which step to take next. The harness gives it the rules, tools and limits for that, and then checks the result.
Does this only apply to software development?
The term comes from software development, but the principle also applies in sales or accounting: set rules and permissions first, then check every result before it goes out.
What does the harness for your team look like?
In 30 free minutes we look at how your team works with AI agents today. Afterwards, you will know which building blocks are missing.
