Why workflows, not agents, are the primitive your team is missing

If your team has adopted AI coding agents, you've probably noticed something strange: writing code stopped being the hard part.

That's the shift Patrick, a principal engineer at Upsun, kept returning to in our latest Product Highlights conversation. He's spent twelve years here, most of them writing back-end APIs in Go, and the past year building with AI on our newest product, Upsun Dispatch™. His verdict on where the bottleneck moved is blunt: "The code isn't really a problem anymore. We can write that really fast. Everything else is still a bit behind the code."

It's the same thing we keep hearing from engineering leaders. AI helped your engineers ship more code. It didn't help your team ship more product. And for a mature team, code was never the bottleneck in the first place.

The bottleneck moved, but the ceremonies didn't

Everyone on a modern engineering team now has an AI agent, often several: Copilot, Cursor, Codex, and a long tail beyond. We stopped counting around 22 major ones, and the list changes weekly.

What hasn't changed is everything wrapped around the code. Patrick calls them "rituals and ceremonies," the same repetitive steps we've always had. You open a merge request. You find someone to review it. You check which requests are clean, which are still under review, and which have conflicts. Multiply that across a team, and it adds up fast, and none of it is the actual coding.

Some of this is already being handed off. Upsun runs an internal review agent tuned to look over merge requests, and Patrick has gone further, asking an agent to triage requests in bulk so he can see their state at a glance. 

The same pattern shows up in design and planning, not just at the code level. As Patrick put it, "We used to think about what can be automated by a computer. And these days, it's pretty much everything."

The risk is that AI is an amplifier. It accelerates good practices and bad ones alike. Point a dozen agents at an unstructured process, and you don't get order; you get faster chaos: review queues growing quicker than anyone can validate, spend nobody can predict, and production systems the team can no longer fully explain.

Workflows are the unit, not the agent

Workflows aren't new to software. What's new is what they can hold. They used to be mostly a tracking mechanism, a way to watch work that humans did by hand. The fiddly, repetitive steps in between had too many variables to hard-code into anything deterministic, so a person did them. AI changes that math. Those steps can now run end to end, with humans staying where judgment is actually needed and stepping out of the parts where it isn't.

This is the core bet behind Upsun Dispatch™: the workflow is the unit of coordination, not the individual agent. Agents come and go. Workflows are how teams get things done. It takes real effort to identify what your workflows even are, then to orchestrate them efficiently and securely. 

Right now they live scattered across agent skills, one-off scripts, and CI jobs, and once someone builds a good one, the whole team wants it.

What that looks like when a team actually does it

You can see the shape of the idea in how we're building Upsun Dispatch™. The team works in one shared monorepo with shared documentation, and that documentation isn't just code. It includes planning, the long-term vision, and the context that usually lives siloed inside marketing, product, and engineering. Everyone has access to the same things and the same vocabulary, and so do the agents.

That last point matters more than it sounds. Patrick is, in his words, "quite pedantic with our glossary," because terms like activity, task, job, and execution have to line up across the stack. One person using a word loosely, and you can end up far down a design path that was off course from the start, for people and agents alike.

The payoff is fewer silos and less waiting. A small question that used to mean holding for a meeting is now answerable on the spot, by a teammate or by an agent that can find the context and act on it. 

It's the same principle Upsun Dispatch is built around: shared context lives in the cloud instead of scattered across individual laptops, so the whole team and its agents work from one source of truth. We already share review prompts and structures between our internal review agent and Upsun Dispatch's. As Patrick says, the way to build the product properly is to eat your own dog food.

What Upsun Dispatch™ actually is

Upsun Dispatch is a platform for governed collaboration between AI agents and cross-functional human teams. The IDE agents gave you coding assistants; none gave your organization a way to govern agents across the board. 

Cursor, Copilot, Claude Code, and Devin each govern their own agents inside their own products. Upsun Dispatch is the layer across all of them, so your team stays in control on one shared platform. Keep your tools and your models; bring them to a common workflow.

Underneath, it comes down to four building blocks. Workflows are structured sequences of agent and human steps that can branch and run on a schedule. Sandboxes keep agents isolated, so nothing touches live infrastructure unless it's meant to. Human gates let engineering, product, design, and security each sign off at the points that belong to them. And an audit trail captures every prompt, diff, approval, and cost, so you have the full record of how a decision got made.

A typical run looks like this: a trigger fires (a Git event, a chat mention, a schedule, an API call) and starts an agent in a sandbox to plan, draft, and execute. A human gate lets the right people decide. The change ships to production, and the whole thing is logged. 

Smart routing picks a model per task for cost and quality rather than locking you to one provider, and because every dollar is attributed to the workflow that spent it, engineering can predict spend, and finance can budget. Spec and Code agents run first, with review, test, maintain, and security agents across the rest of the lifecycle.

Who's building it

Upsun has spent over a decade responsible for other people's production environments at scale, and Upsun Dispatch applies those same standards to agent workflows. 

Leading the effort is Fabien Potencier, creator of the Symfony framework and Upsun's Chief Technology and Product Officer. He spent 20 years solving the large-scale open-source software lifecycle; Upsun Dispatch is his take on the agentic one. If the maturity question interests you, his write-up on the 8 stages of AI engineering maturity, adapted from Steve Yegge's levels of AI-assisted development, is the best place to start.

The Upsun Dispatch design partner cohort is open now. 

Patrick's reason for building it is refreshingly self-interested, and probably the best pitch there is: “I'm interested in Upsun Dispatch to make my own life easier. And that's hopefully how it's gonna work for everybody else too.”