Don't build the autonomous AI factory first

Here's a scene playing out in engineering teams right now. An engineer spends the weekend running four or five coding agents in parallel. Monday morning, a teammate opens their laptop to 53 changed files with 2000+ diffs and a message that says, more or less, "should be good to merge." Nobody asked for this much output. Nobody has time to review it properly. The team doesn't feel faster. It feels ambushed.

That's what happens when individual AI productivity outruns team-level readiness, and it's the clearest complaint we heard across conversations with the people building Upsun Dispatch™. This is why AI adoption has to happen in order, one stage at a time, and Fabien Potencier's eight-stage maturity model is the clearest map of what that order actually looks like.

Teams that jump from the vacuum straight to the autonomous factory end up exhausted, exposed, or both. The ones that get real value go through every stage in between. Here's why that order matters.

The stages are a sequence of trust problems.

Fabien Potencier, Upsun's CTPO and the brain behind Symfony, put it simply: you have to live through each stage of AI adoption to understand why the next one exists. 

His model runs from the vacuum (stage 1), where AI use has no real strategy behind it, through the drift, the islands, the standardization bet, the workflow redesign, and the operating system, to the bright factory (stage 7), where engineers supervise a growing swarm of agents running locally, and finally the autonomous factory (stage 8), where agents run on shared infrastructure and the job becomes managing the system rather than the code.

Each stage isn't just "more AI." It's a different problem: how much you trust the output, how exposed your systems are, and how well the work holds together once more than one person is touching it. You don't solve the standardization bet (stage 4), where leadership commits to shared tooling and governance, by installing that tooling on top of a team still stuck in the drift (stage 2), where one person's workflow only lives in their own head. The tool doesn't do the work of building the trust. The team does, by going through the stages that build it.

The Monday-morning scene above is what the bright factory (stage 7) looks like when nobody built the habits the earlier stages were supposed to teach.

What breaks when you skip ahead

We heard the same three failure patterns from multiple people, independently, which is usually a sign you've found something real rather than one person's pet theory.

1. Security debt piles up quietly. Patrick Dawkins, Upsun’s Principal Software Engineer, calls the core risk of unsupervised agents a "lethal trifecta": untrusted data, real actions, and private information, all in reach at once. Teams catch this early, reviewing every command. The catch fades as approvals get repetitive and people start clicking through unread, well before anyone decides to turn it off.
2. Technical debt compounds faster than teams expect. It's a pattern that came up more than once in our conversations: a project that used to accumulate five years of debt before it needed a rewrite now hits that point in six months.
3. One person's productivity becomes everyone else's bottleneck. This is the Monday-morning scenario above. The gains from running agents are individual. A single engineer can genuinely do more. But hand that same engineer's output to a team of three or four that hasn't built a shared process for reviewing agent-generated work, and the team's overall output can stay flat or drop. 

Autonomy is earned, not switched on

If skipping stages is the failure mode, the trust dial is the mechanism that prevents it. Agents earn autonomy the same way a new hire does. At the start, every meaningful action requires a human to check it. 

Over weeks, as a specific workflow proves itself correct again and again, the human gate on that workflow gets lifted, and the agent moves on to something less closely supervised. Trust isn't granted up front. It's earned separately for each workflow rather than all at once for "AI" in general.

Models are currently just slow enough that a human can follow along, but that won't last. The point of staged trust was never to keep humans reading forever. It's to make sure the autonomy that eventually replaces that reading was actually earned, not skipped.

The stages apply no matter where a team starts

It would be easy to read this as "you need a mature team before AI is worth using." That's not the argument. Teams in the vacuum (stage 1) can and should begin. What doesn't work is skipping the stages that build the muscle a later one assumes you already have.

Starting in the vacuum on purpose beats parachuting into the standardization bet (stage 4) and hoping the workflows show up on their own. If your team is still there, start on purpose, not because you haven't reached the workflow redesign (stage 5) yet, but because the vacuum is where you build the review habits that stage assumes you already have.

Proof from the team building Upsun Dispatch™

Upsun Dispatch was built using an early version of itself, and that early version caught exactly the kind of problems staged adoption is supposed to catch. The first version of the pull request review workflow dumped every piece of feedback into one large comment. 

It took actually running the workflow, not just designing it on paper, to see that this made reviews harder to use, not easier. It was rebuilt to leave comments inline, next to the code they applied to. This staged rollout doesn't just prevent disasters; it surfaces the ordinary mistakes that only show up once real work is running through the system.

Where this leaves you

None of this says wait. It says go in order.

If you're already running agents in parallel and it's starting to feel like the Monday-morning scenario above, that's not a sign to add more automation. It's a sign to go back a stage and build the shared workflow you skipped.

The teams that get this right won't be the ones who reached the autonomous factory first. They'll be the ones who can point to every stage they went through to earn it.

For the fuller picture, Fabien Potencier mapped all eight stages, from individual habit to organizational habit, in a separate post.

• New here? Read the introduction to Upsun Dispatch.
• Want to help build it? Apply to the founding design partner cohort.