I help teams get real value out of their coding agents
I am Theo Martin, former Tech Lead of the Core AI team at Akeneo. I work with teams that use a coding agent, a tool like Claude Code (Anthropic’s), Codex or Cursor that reads, edits and runs a project’s code on request. Developers, and also product managers and designers, for whom the agent becomes relevant in turn.
Three situations where people call me
The agent fails
It gets your tickets wrong and you do not know why. On your own tasks I look at what it gets right, what it gets wrong and what makes the difference: context that is written nowhere, no way to check the result, an instruction that holds it back. Two tasks that look close to us can be far apart for it, and the map of what it can do is drawn on your tickets, not on an example.
Adoption does not take
You are a lead or a CTO, the tools are paid for and handed out, and your developers use them little or not at all. It is almost a diagnosis of behavior: a first failed try that convinced people it does not work, a first win that convinced them they had it figured out, a developer identity, a tool banned for a long time. The mix differs from one team to the next, I first look for which one is yours, then I convince your developers on their own code.
Learning to use it better
You already use it and want to go further. We start from a mental model of what an LLM is, what it knows and what nobody wrote down for it (it was not at the stand-up), so it can be wired to Jira, Confluence or Slack and left to look for its own context instead of being fed copy-pasted. It serves developers, and just as much PMs and designers.
In late August 2026 I read 269 public repos file by file to see who owns the files that steer their agents. PostHog does not use CODEOWNERS (the GitHub file that says who must review what) for them: ownership is declared in an owners.yaml file that blocks no merge. Elastic, on Kibana, does the opposite and makes its tech leads owners of /AGENTS.md and /.claude/, so their review is requested on every change.
What I do
Nine years of ML and LLM systems in production, three and a half of them at Amazon on European pricing. Then Tech Lead of the Core AI team at Akeneo, a B2B SaaS, where the inference platform I architected served every product team and, behind them, hundreds of enterprise retailers and tens of millions of product attributes generated every month.
That is where I got my team onto code agents as soon as they shipped, gave classes to several teams, and convinced people who then convinced their own teams in turn. It is also where I shifted from the tool onto the harness (the set of instructions, tools and checks you put around the model), with CLAUDE.md files, the instruction files the agent reads, kept in the repo and reviewed like code. I use agents every day, at a rate of hundreds of billions of tokens.
I have worked inside the repos of large companies, one of them in the CAC 40: real enterprise repos, not demos.
The NDA means I say nothing about the client, its sector, or its stack. The method I can describe in full, the artifacts I cannot, and I would rather say that before the call than during it.
Testing rather than believing
One example of what I call testing. A .env file holds a project’s passwords and access keys, you do not want the agent to read it, and Claude Code lets you write a rule that forbids reading it. As of October 5, 2026, that rule covers the agent’s read tool, and also cat, head, tail, sed and redirections, since version 2.1.257 of September 1. It does not cover a grep -r run from the folder, nor a script that opens the file itself. For a real guarantee you need the sandbox, an enclosure the operating system imposes on the agent’s commands, whatever the text of the command says.
I tell the state rather than the rule, because it will have moved by the time you read this. A security rule placed on the text of a command is a help, never a boundary, and it gets retested on your repo at every version.
What the work looks like
What gets written during an engagement is rules kept in the repo and reviewed like code. The CLAUDE.md of this very site is one, in French, readable in full on GitHub. The excerpt below says, roughly, that the build script stays under 300 lines, that three obvious lines beat one clever trick, and that a template marker left unsubstituted fails the build on purpose.
## Modifier le build
`build.py` tient en 300 lignes et se lit de haut en bas. Il n'a qu'une dépendance externe, `markdown`. Entre une astuce élégante et trois lignes évidentes, écrire les trois lignes. Un ajout qui ferait dépasser le budget de lignes est le signe qu'il faut retirer autre chose, pas augmenter le budget.
Un marqueur laissé non substitué dans une page produite fait échouer le build. Inventer un marqueur dans un gabarit sans l'alimenter dans `build.py` casse donc le déploiement, ce qui est le comportement voulu.On my own repo, cutting a CLAUDE.md from 108 to 30 lines gave the same result on my test bench: 22 cases out of 22 before and after, against 2 out of 16 with no file at all. The bench is small and its judge is a simple rule, I say so. Length changes nothing, absence does, and it is the kind of thing you can measure at home in an hour.
Four formats
If you are looking for a catalog Claude Code course, two days in a room on a sample project, that is not what I sell, and I went through the public prices of those who do in a separate piece. I turn up in your repo, on your tickets, and what gets written during the day stays there, the format is on the training page.
Agentic setup audit
€1,500 excl. VAT per day, fixed price, two to four days
I look at what is written for the agent in your repo, what still applies when it switches tools, what blocks a merge and where the bill goes. You end up with a written playbook and a 90-day plan you can run without me.
Workshop in your own repo
€2,500 excl. VAT per day, billed by the day and never per head
One day on your own code and your real tickets, with your developers at the keyboard rather than me.
Office hours
€300 excl. VAT per hour, or €1,400 excl. VAT a month for four sessions
An hour with your leads on whatever is blocking right now, screen shared, end of day included.
Building an agent system
€7,500 excl. VAT per week, two weeks minimum
For when what you need is not an opinion but shipped code, an agent wired into your product, an eval harness that tells you whether the answers are any good, a RAG that holds up under load, an inference platform that does not fall over. I write the code, your developers review the code and the CLAUDE.md that drives the agent, all of it lives in your repo.