
AI Productivity Lead @ Sanofi — scaling AI engineering + measuring real impact in regulated pharma. @AnthropicAI Ambassador • Paris AI meetups host.
Paris
Joined October 2010
- Tweets3.2K
- Following72
- Followers258
- Likes812
Pinned Tweet
Most AI content on this feed is theory. Mine is from the other side.
I scale AI engineering across a global regulated pharma and measure what actually works.
Posting what it takes.
Field reports, unpolished takes, occasional hot takes.
92
1/
Your agents shipped 50 PRs overnight.
Your team just inherited 50 decisions they never asked for.
1
20
5/
The Three Clocks:
→ Machine speed for execution
→ Human speed for decisions
→ Business speed for adoption
Most orgs run all three on one clock.
1
10
Simon Willison puts AI-generated PR descriptions behind a collapsed section. I'd keep the review brief visible: what changed, why, and what check shows it meets the spec. Put the generated detail underneath, not the evidence needed to review it.
xitter.cf/simonw/status/21196503…
I've recently started hiding the bulk of my AI-generated GitHub PR descriptions behind a collapsed <details><summary> HTML tag (thankfully supported by GitHub Markdown) and I think it's a useful pattern for making PR slop a little bit less exhausting simonwillison.net/2026/Nov/9…
22
I'd use @mattpocockuk's /grill-with-docs to recover the architecture decisions that never made it into the docs:
/grill-with-docs on our existing architecture, not a new feature: grill me only about decisions that aren't written down anywhere. Every answer lands in GLOSSARY.md or an ADR.
Stop when a new teammate could re-derive our design from the docs alone
20
Finding vulnerabilities got cheap. Reading the patch and checking nothing else broke did not. In our team the bottleneck moved from writing code to reviewing it, and review doesn't scale with a model.
12
Almost all the code in the 200-person org I work with is now AI-assisted. The surprise: writing it isn't where we got stuck. Writing the spec, and the check that proves the agent did what we meant, is. Nobody taught us that. We're learning it in production.
19
Unpopular take: AI doesn't replace specialists, it replaces average specialists.
The top 1% get a lever. The competent, certified, usually-right middle now competes with something that works at 3am and never gets bored.
That's the part nobody puts on a slide.
13
The spec is the new code.
Writing the code used to be the bottleneck. Now an agent writes it in minutes.
What is left is saying precisely what you want, and proving you got it.
Intent in, validation out. That is the engineer's job now.
11
Agents are becoming the experts.
The single-field specialist now competes with an agent that never sleeps.
The generalist directs ten of them.
The scarce skill is no longer knowing the answer. It is knowing which question to ask.
12
Thomas Roux retweeted
Exactly a year ago, I launched Beads, my agentic memory layer, issue tracker, and knowledge graph. I built a sketch of it in fifteen minutes with Claude Sonnet 4.5, and the rest was history. Today Beads has a sophisticated, scalable backend, a strong contrib team, and 1.5M downloads.
You may have your own favorite tools, but Beads is every agent's favorite tool. Show it to your agent today.
blog.gascity.com/posts/fun-f…
Thomas Roux retweeted
a couple months back i was on a pod with @vtahowe and @ianlivingstone chatting w/ a Billions-of-market-cap CISO about how to balance giving developers tools to make coding agents useful (think, mcps to all kinds of useful context link ticketing, email, slack etc) while avoiding data exfiltration, prompt injection, credential hijacks, etc. There was some hand-wavy "mcp gateway" talk but really cool to see this important question materialized into a formal spec for builders.
There's a lot more to build here, but this is a great start
Today we're publishing the first draft of the Agent Baseline Software Factory Reference Architecture!
An opinionated blueprint for securing increasingly autonomous software factories, built with folks across the community.
It's open for feedback and contributions!
agentbaseline.org/reference-…
In 2012, Michel Serres saw Petite Poucette challenge authority: knowledge was now in everyone’s pocket. AI takes this further. As agents take on deep expertise, our value shifts to framing problems, judging outputs and connecting ideas. Authority comes from helping others grow.
62
One thing to test today? @humanlayer_dev
Don't know where to start your day? Start here.
RPI workflows, human-in-the-loop on the design doc, no 2,000-line surprises. This is what AI-native engineering actually looks like.
Great work @dexhorthy & @0xblacklight
2
8
850







