The longer, personal version of this is the essay about the AI teammate a whole company ended up using. This is the short, forward-looking one: what I think it means, and what I want to do about it.
The bet
The next decade of engineering teams will run on autonomous agents the way this one runs on version control: unremarkable, load-bearing, and impossible to work without. Not as a feature bolted onto a product, but as teammates that carry context, do the routine work, and give people back the hours only people should be spending. The teams that adopt this well will move at a speed the rest simply cannot match.
Why now
Models have finally crossed the line from impressive demo to reliable colleague on bounded, well-scoped work. That is the unlock. It is also where most people stop, because they assume the hard part is the model. It is not. The hard part starts the day you put an agent in front of people who have real work and no patience for a tool that might embarrass them.
What's broken
Almost every org that tries this makes the same mistake: they treat adoption as a technical rollout. Ship the capability, announce it, expect uptake. It does not work. Agents do not get adopted because they are smart. They get adopted because a skeptical person decided to trust one, and trust is earned team by team, week by week, in a currency nobody's tooling measures.
So the real gap is not capability. It is the entire layer between "the agent works" and "the org relies on it": trust, legibility, humans in the loop by default, and the ability for every team to shape an agent to how they actually work. That layer is unglamorous, it is mostly not about models, and it is exactly where the durable company gets built.
What I think the winning shape looks like
- Trust-led, not capability-led. The product is the thing people stop worrying about, not the smartest thing in the room.
- Multi-persona from day one. One team's "helpful" is another team's "won't stop talking." The platform bends to the team, not the other way around.
- Humans in the loop where it counts, by design, not as an apology. Control is what makes autonomy acceptable.
- Boring on purpose. The wins that compound are the routine ones, done reliably, thousands of times. Flashy wins a demo; boring wins a habit.
None of that is a model problem. All of it is a product-and-trust problem, which is precisely why I think it is worth spending the next chapter on.
Where I am
Early, and building. If you are a founder chasing the same thing, an operator who has felt this pain from the inside, or an investor who thinks hard about this space, I would like to talk. Email me.