Skip to content

Freelance in France · remote

Harness engineering for AI coding agents

Harness engineering covers everything that surrounds the model of an AI coding agent: the instructions it reads, such as AGENTS.md or CLAUDE.md, the tools and MCP servers it can call, hooks and skills, the sandbox it runs in, the tests and the validation loop in CI.

I set it up with your team, in your repositories, with the tools it uses: Claude Code, Codex or Gemini. The agents work on your real tasks, CI checks every commit, and the team decides what gets merged.

What the engagement covers

Each part is written in your repositories, versioned with the code and documented.

Instructions for the agents

An AGENTS.md or CLAUDE.md file describing the project’s architecture, conventions and commands, kept up to date with the code.

Tools and MCP servers

The tools the agent can call, and the MCP (Model Context Protocol) servers that give it access to your systems, each with defined permissions.

Hooks and skills

Hooks that block a forbidden action at the moment the agent attempts it, and skills that describe the team’s recurring tasks.

Execution sandbox

An isolated, reproducible environment where the agent installs the project, runs the code and the tests without access to production.

Tests and the validation loop in CI

Reliable tests and architecture checks on every commit and in CI. When a check fails, the agent gets the error and fixes it before any human review.

Context for each task

What the agent receives for a given task: the relevant files, the decisions already made, the done criteria. This work, called context engineering, is part of harness engineering.

How an engagement runs

The engagement starts with a small scope: one repository, a few real tasks, and written rules for the agent from day one.

Then come the automated checks in CI, set to your standards, your tests and your architecture. Each gap found on the following tasks becomes a rule or a check.

Once the frame has proven itself, it extends to the other repositories and the rest of the team. The same path works with Claude Code, Codex or Gemini.

Agent readiness

Before extending the use of agents, I check observable criteria in the repository and in how the team works:

  • reliable tests, which give the same result on every run;
  • a reproducible setup, which an agent or a new developer can run without help;
  • an up-to-date instruction file for the agents;
  • clear boundaries: what the agent may change, and what needs human approval;
  • a CI that blocks the merge when a check fails;
  • code review that keeps pace with contributions.

This check can be done on its own, as an agent readiness audit: you get a written assessment with the priorities in order.

Reviewing code produced by agents

When agents produce more code, review becomes the point where work piles up. Automated checks take over what can be verified mechanically: conventions, tests, dependencies between modules. Human review focuses on the choices that need judgement.

This practice is sometimes called agentic engineering: the developer specifies the task, reviews the diffs, checks the result and stays responsible for the merged code.

With the team, I write the review rules: what an agent may propose, what a human must review, what CI rejects before any review.

Questions about harness engineering

What is harness engineering?

It is building the working environment of AI coding agents: the context they receive, the rules they follow and the automated checks that validate their work. The code the agents produce stays verifiable, maintainable and true to the project’s architecture.

How do you speed a team up with AI agents without losing quality?

By automating what review alone can no longer cover: every commit and every merge request goes through checks on your standards, your tests and your architecture. Human review focuses on the choices that need judgement, and the agents produce more without quality dropping. This is what the harness engineering offer covers.

Where do we start with Claude Code in a team?

With a small scope: one repository, a few real tasks, and written rules for the agent from day one. Then come the automated checks in CI, and the move to other repositories once the frame has proven itself. The same path works with Codex or Gemini.

Let’s talk about your agents and your CI

A few lines about your repository, your tools and what is getting in the way are enough. Scope gets clarified in an initial conversation.

Go to the contact form