16 years writing code. Now I make sure the machine writes it well.
I'm Flavien Beninca. I make AI agents productive inside large codebases: the architecture they inherit, the context they get, and the tooling a whole team needs to do it consistently. My job used to be writing the code. Today it is designing the rails that decide whether an agent ships something you would keep — or something you will rewrite in three months.
What I am building
Getting one agent to behave is a solved problem. Getting a whole engineering org to do it consistently — and knowing whether it worked — is the actual job.
- Turnkey agent config, per role
- A tool that distributes Claude configuration — skills, settings, MCP servers — tailored to who you are. A product manager does not need a frontend developer's skills, and a backend developer does not need the same set as the frontend one sitting next to them. Everyone gets a working setup instead of a wiki page they will not read.
- Telemetry on agent usage
- You cannot optimise what you do not measure. This one tracks token and context consumption over time, so we can answer real questions: did that optimisation actually help, which models are quietly expensive, and are the skills and MCP servers we shipped being used at all.
- An MCP gateway
- Before the config tool existed, sharing MCP servers across teams was the bottleneck. Putting GitLab, Jira, Figma and the rest behind a single gateway made distribution a non-problem — one endpoint to hand out instead of one setup per person, per server.
What it rests on
None of it works without the basics. Four things decide whether an agent ships something you keep.
- Architecture
- An agent inherits your structure. A codebase full of god modules and implicit coupling generates bad code for the same reason it always did — there is just a lot more of it, a lot faster.
- Context engineering
- Explicit, reviewable rules the tools actually read, shared across the team. Rails, not prompts. Context is also a finite resource: most agent failures I see are a polluted context, not a weak model.
- Subagents
- Splitting work so each agent gets a narrow, clean context, instead of one long session that slowly drowns in its own history.
- Review
- The bottleneck moved from writing to reviewing. AI finds patterns, humans confirm intent — and the engineering judgment now lives in the review, not the keystrokes.
What it buys
22s → 4s
Test startup per file, after an agent found the architectural flaw instead of me tuning config flags.
Read the write-up →15 min
Saved on every CI run, from that same refactor — roughly 90 files updated by the agent.
Read the write-up →2900+ tests
Migrated from Jest to Vitest on a React/TypeScript codebase, ESM handling included.
Read the write-up →0 lines
Written by hand on a side business now live in production — the run that convinced me the job had changed from typing to orchestrating.
Read the write-up →The path
- Started fullstack — PHP, MySQL, and the servers under the desk. Then a decade across construction, real estate, e-learning, delivery and public administration.
- Still fullstack — CakePHP on the back — but Vue.js is where the pull started: a learning management system, a serious game shipped as a desktop app, an internal CRM.
- Lead frontend developer in a four-person agency. React and Apollo over PHP and Node microservices — including a data platform for a French public agency, released open source.
- Lead frontend, full remote, on a construction-industry SaaS. Hired to take a Symfony monolith to an independent React front end, alongside the design system and a team of nine frontend developers. TypeScript, linting, unit tests with coverage, real reviews, CI.
- New role, same company: making agents productive across the whole engineering org — not just in my own terminal.
I do not work for a company that sells AI products, and I do not sell training. What is on this blog is what I actually run — including the parts that break, loop, or turn out to be a waste of tokens.
Get in touch
If you are working on the same problem — making a large codebase safe for agents to work in — I am interested in the conversation.