The clearest sign we had built something real did not show up in a dashboard. It was teams I had never worked with, messaging out of nowhere to ask how they could get one too.
That was the moment I realized I had spent weeks being proud of the wrong part of it. I had been proud of the agent. What mattered was the pull.
We built it, originally, for one team drowning in the kind of work that is too small to prioritize and too frequent to ignore. An agent that could pick up the routine tasks, carry context from one to the next, and quietly get them done. We did not announce it. We did not pitch it. A few weeks later a second team asked for one. Then a third. Then teams I had never met.
Here is what I did not understand going in. We had not shipped a piece of software. We had hired a colleague nobody interviewed, and now every team in the company had to decide, on their own, whether to trust it.
Everything interesting that happened next was about that decision.
What I’m not going to tell you
I’ll be honest up front: I’m not going to explain how it works. Not the architecture, not the models, not the clever bits. Partly because the clever bits are the least interesting part of this story, and partly because the how is a longer story for another day.
So this is the other half. Not how we built an agent, but what it taught me about putting one in front of busy people who have real work and very little patience.
Trust is the product
Every team’s first question was a version of the same one: “what happens when it’s wrong?” Never “how smart is it.” They had all seen the demos. What they wanted to know was whether this thing would do something confidently stupid with their name on it.
So the real work, once the agent basically functioned, was not making it more capable. It was making it trustworthy, one week of real behavior at a time. Trust does not install. It accrues, or it leaks, and you rarely get to pick which.
Capability got us to the first demo. Trust got us to the fifteenth team.
Boring beats flashy
The tasks that won people over were never the impressive ones. They were the tedious ones. The chores too small to staff and too constant to escape. The work that eats an afternoon and leaves nothing to show for it.
Flashy wins a demo. Boring wins a habit. And a habit is the only thing that survives the week after the novelty wears off.
By the time it had settled in, the agent had raised and merged more than 2,000 pull requests, been onboarded by over 15 teams, and touched more than 100 codebases, giving people back hundreds of hours they used to lose to the dull stuff. None of those numbers came from a moonshot. They came from doing the small thing, reliably, thousands of times.
Every team wanted a different colleague
I assumed one good agent would serve everyone. It did not.
One team’s “helpful” is another team’s “won’t stop talking.” One wanted it proactive. Another wanted it silent until asked. The same behavior that made one team love it made another quietly switch it off. A colleague who acts identically in every room is not polished. They are oblivious.
The platform had to bend to each team, not the other way around. The day we stopped shipping one agent and started letting each team shape their own, adoption stopped being something we pushed and became something they did without us.
Keep the human where it counts
The fastest way to lose a team is one confident mistake on something that mattered.
So the agent asks. On anything with real consequences, a person stays in the loop and signs off. I worried this would feel like training wheels. It did the opposite. The approval step was not the tax people tolerated. It was the reason they let it in at all.
Autonomy is not the absence of a human. It is a human who has learned they can look away.
The bottleneck was never the model
If I had to compress the whole thing into one line: adoption is a people problem wearing a technical costume.
We spent far less effort making the agent smarter than making it legible, predictable, and easy to say yes to. The hard questions were never “can it do the task.” They were “will a busy, skeptical person believe it, adopt it, and keep it.” Those are questions about humans, and there is no model you can swap in to answer them.
What it costs
None of this is free, and I won’t pretend otherwise. High-stakes actions still need a person. Quality depends on measuring the thing relentlessly, which is a discipline of its own. And every new team is real work, because every new team is a fresh set of expectations to earn from zero.
But the shape of it is hard to argue with. Something built for one team got adopted across a company, not because anyone mandated it, but because it made people’s days better and they trusted it enough to let it in. I do not know a more honest signal of demand than that.
It is also the most convincing thing I have seen about where this is all going, and it is what I want to spend my next chapter on.
If you are building one of these too, the model is the easy part. Spend your time earning trust, not adding features.
The teammate who gets kept is not the smartest one in the room. It is the one people stopped worrying about.