Just 10 years ago, I was privileged to join the small early team entrusted with laying the foundations of Michelin's digital transformation.
We were small, but the ambition was not. What we were building was not yet visible either; it was infrastructure, capability, culture, and trust more than it was products anyone could point to. Very early on, a few convictions started to guide us: build from within, internalize talent and delivery, master enough of the stack to remain sovereign, and invest in foundations strong enough to last. Looking back, I realize that many of those choices were less about technology than about agency.
A Team Grown Like Software, Operating Like an Agent
For a long time, I liked to think of our team as a piece of software. The analogy made sense: a distributed organization made of several hundred people, operating like a true cloud of talent, continuously evolving through iterations, migrations, refactorings, and a few painful but necessary rewrites. It also felt increasingly composable, with empowerment at the edge and interfaces turned both inward toward the Group and outward toward a broader ecosystem, including the valuable perspective that freelancers have long brought to our teams.
In short, these are largely the traits of modern software itself. And this is no coincidence, because as Notion's founder Ivan Zhao liked to say, "First we shape our tools, thereafter our tools shape us."
But over the past months, as our teams have embraced agentic AI more fully, whether in software delivery itself or in individual productivity, I have started to feel that this comparison no longer goes far enough. Our teams do not really feel like software anymore. They feel more like an agent.
Not an AI agent in the fashionable sense, but an agent in the more useful sense: a system with agency—that is, the ability to act, to decide, and to move without waiting. And the more I looked at it, the more the parallel held.
So let me take this analogy seriously for a moment, and follow the traits of a good AI agent one by one: skills, tools, workspace, memory, learning, and protected autonomy. What strikes me is how well this vocabulary ends up describing the principles that have guided our team over the past ten years, and perhaps why this metaphor suddenly feels less playful than clarifying.
Skills
A good agent has skills: it can execute.
Over time, our teams became less a structure that merely "supports digital" than teams able to design, build, and run across product, engineering, platform, design, content, and now AI-native forms of work. Skills are what turn intention into execution, and vision into something tangible enough to matter.
Over the years, this has translated into very concrete outcomes: rebuilding more than 500 websites and mobile applications across the Group, redesigning their UX through a global design system, internalizing marketing intelligence at scale, and platformizing Michelin's geodata capabilities. It also meant fostering innovation much more deliberately, through digital partnerships ranging from startups to the largest tech players, through countless hackathons and learning expeditions that helped widen our collective field of view. In reality, we have done much more than that. But what matters to me is that these achievements were never just a portfolio of deliveries. They were signs that a different operating model was emerging.
Tools
A good agent has tools: it interacts with the world and sometimes even builds new tools for itself.
This has always felt deeply familiar. A large part of our story has been about refusing the false choice between consuming the off-the-shelf tools of others, or trying to build everything ourselves. What we tried to do instead was to master enough of the stack to choose wisely, assemble deliberately, and build what needed to exist when the surrounding tooling was not enough.
Open source has been a major marker of that conviction. Not only because it often gives better building blocks, but because it tends to keep technology discussable, inspectable, and less captive. Over these ten years, we have built on the shoulders of giants, used some extraordinary technologies, contributed back a little ourselves, and occasionally uncovered real gems along the way. We have long believed in open ecosystems, in platforms attractive from the outside as much as from the inside, and in the kind of technical culture that seems to grow stronger when it stays connected to communities larger than itself.
Workspace
A good agent has a workspace: it does not operate in theory, but in context.
For me, the workspace is first the field of work an agent is allowed to act upon. The larger and better defined that field becomes, the more room there is for action, impact, scale effects, and the spread of patterns. In many ways, this is also what happened to our teams over the years: the scope of what we could work on kept expanding, from websites and apps to platforms, geodata, marketing intelligence, content, AI-assisted delivery, and more. Of course, this only works within the limits of actual capabilities and resources. Too much context kills context. But without a meaningful workspace, there is no meaningful agency either.
Filesystem
An agent's filesystem is not just storage; it is the accumulated patrimony of what has been produced along the way.
The same is true for teams. In an internalized platform model, you do not simply produce the result; you also build the factory at the same time. The filesystem, in that sense, is both the output and all the production artifacts that make the next output easier, faster, more coherent, or more reusable. Seen from the outside, it may even look messy, opaque, or hard to read. That does not matter much. What matters is that it remains usable, alive, and keeps creating value for those who work inside it. Some of the most important things we built over the past ten years were not always visible from the outside: platforms, APIs, design systems, shared foundations, and bits of invisible renovation work that made later acceleration possible. The best teams do not just deliver outputs; they also leave behind assets that change what can be produced next.
And beneath all of this sat culture. We grew, from almost nothing, into a team of more than 450 people, multiplying our scale by roughly a hundred over the decade. That kind of growth does not happen through org charts alone. It requires a culture strong enough to pass on standards, trust, ambition, and care at scale, without weakening initiative. If there is one thing I am especially proud of, it is that beyond the sites, apps, platforms, and systems, we also helped grow a human environment in which those things could keep emerging.
Sandbox Is Not a Bug, It's a Feature
That is where this analogy becomes more than a metaphor. Teams trying to build new digital capabilities inside large organizations often need a certain distance from default tools, habits, and operating patterns. Not as a posture, and certainly not as a claim to exception, but simply because some forms of learning, building, and structuring need room before they can be shared more broadly.
For teams as for an agent, isolation—or rather a protected space, a sandbox—can sometimes be a feature rather than a bug.
An agent without a sandbox is not more powerful; it is more brittle, too exposed, too interruptible, and too dependent on the ambient environment to hold any coherent course of action. I think the same is true for teams trying to create real digital capability inside large organizations. They need interfaces with the broader system, shared outcomes, interoperability, and strategic alignment. But they also need a workspace of their own, with enough room to assemble their tools, develop their skills, enrich their context, and learn by doing, not in opposition to the rest of the organization but in service of it.
This can also lead, at times, to a setup that differs somewhat from the rest of the Group's, whether on the software or the hardware side, simply because strategic autonomy sometimes requires a slightly different workspace.
That is also why I have always been cautious with purely top-down standardization. At scale, standardization matters, of course, but imposed too early, or too uniformly, it often creates compliance without capability—I myself experienced it first-hand numerous times when welcoming diverse teams in our org. The most resilient systems usually know how to preserve differentiated spaces of experimentation, as long as they remain legible and connected to the whole. That, to me, is the real promise of a sandbox: not separation for its own sake, but autonomy that remains useful to the whole.
Learning
This matters even more now that the language of software is quietly giving way to the language of agency.
After years of relative stability in digital ecosystems—however complex they remained from an economic and strategic standpoint—it feels as if we have entered a supercycle of disruptions that is re-challenging almost every assumption on which the previous decade of digital transformation was built.
Interfaces are collapsing back into conversation, CI/CD is being rethought in the age of AI, much of the SaaS market seems to be pivoting under our feet, and individual productivity tools are turning into agentic environments rather than simple applications. Our traditional categories of tech profiles are blurring as well. And content itself, from marketing assets to pure code, is becoming increasingly liquid and recomposable, shifting more and more value toward the layers that can still be architected. Knowledge work, in turn, is becoming more contextual, more tool-augmented, and more workspace-native. In that world, the teams that may prove most useful are not necessarily the ones that simply consume the approved stack; they may be the ones that know how to shape their tools, grow their skills, curate their context, and create working environments where people can move fast without losing coherence.
Learning also comes from de-centering, from the sidestep, from exposure to edges rather than cores. We first lived that in Lyon, in contact with agile and frugal teams operating at the periphery of our very large Group. We lived it through the constant and massive presence of expert freelancers from many horizons, who brought a valuable outside perspective while often developing a deep sense of belonging to the Michelin family. We lived it again through the major extension of the team in India seven years ago—and I still feel that personally now that I have moved there myself, in what has become a fully distributed team in the most concrete sense. And we lived it through the many reorganizations and recompositions that kept mixing people, practices, and ideas, always in close contact with external ecosystems.
Memory
A good agent has memory: it knows what matters.
Perhaps this is why I wanted to keep memory for the end of this analogy. At the very moment we are celebrating the ten years of a team that started from almost nothing, looking back at what these years have contained also helps illuminate what comes next. Yes, almost everything has changed on the technological front, and almost everything will keep changing.
And the most vivid memory of all is probably the memory of faces: all those team members whose presence brought us joy over these ten years, in France and in India, through sleepless production nights, bursts of laughter, afterworks, and all the intense moments in between. That, too, was the forge in which a shared identity slowly took shape.
But collective memory is also what prevents movement from becoming drift: what we have built together, what we have learned, what we have defended, what we have had to unlearn, and what convictions proved worth carrying forward. In a technological and business ecosystem that only grows more uncertain, this kind of memory does not make us nostalgic. It helps us know what to keep, what to change, and what to resist.
In the end, from skills to memory, from workspace to autonomy, this analogy has helped me name something I had mostly felt until now. What best describes this team is not software moving from version to version, but a system of agency: one that acts, learns, builds, leaves behind more context than it found, and turns intent into value, again and again.
The Bicycle Lesson
Now what about intent? A good agent does nothing without a prompt. But in the case of teams, the best move rarely comes from a top-down prompt or a single instruction handed down from above. It is something looser, and in a way more powerful: a shared orientation, an energy, a common pull.
Recently, Philippe Jacquin reminded us of a deeply inspiring analogy from Michelin's earliest days: the bicycle—which is no trivial image for a company whose story began, quite literally, with bicycle tyres. A bicycle has this unique property that it only stands if it moves; it evolves in one direction only—forward—and it never quite moves in a perfectly straight line. Its constant motion, its subtle imbalance, even its winding trajectory are not weaknesses. They are precisely what makes balance possible.
This lesson resonates deeply with everything that has animated me over the past ten years. Digital teams do not stay alive by freezing into a final form. They endure by moving, adjusting, learning, and finding their balance in motion. This team was not designed from above. Together, we have grown it over the years.
This idea also resonates with the recent and inspiring book Le commandement ne dort jamais—Transformer ou périr, a collective work led by Emmanuelle Duez. It follows the insomnia of a fictional leader confronting the transformation of his own organization, enriched by the crossed perspectives of its co-authors. One passage stayed with me in particular:
"To be marginal is not to stand outside the system, but at its frontier, where it still breathes [...] one foot inside the frame, another outside."
And perhaps many teams tasked with transformation build their legitimacy that way too, with one foot inside the frame and another outside.
So as I look back on these ten years, I feel above all deeply grateful to those who made all of this possible. First, to Eric Chaniot and Patrice Cochin, who founded this team and gave the Lyon "sandbox" the space to exist and grow, but also the ambition, the challenge, and the first structure that allowed it to become real. To Yves Caseau, for allowing digital factories to scale within the broader Michelin IS/IT organization, to the dimensions of the Group. And for consistently giving the team what it needed next: stronger business orientation, a forward-looking vision of software engineering, and the confidence to keep turning intuition into contribution.
And finally, and most importantly, to my exceptional teammates: Benoit Lamouche, Victor Rassion, Hugo Mourlevat, Florence Lecuyer, Stephane Maccari, and Xiaojing Chen: every day, you lead with passion, expertise, and generosity, together with the wonderful teams around you. Whatever agency this team has built over the years, it lives first and foremost through you.
