Every software process we've invented is, in some way, a response to human limitations. Sprints exist because we can't sustain focus indefinitely. Standups exist because we can't see what's in each other's heads. Tickets, handoffs, code review, pull requests—a lot of our development process is scaffolding around what people forget, how much context we can hold at once, and how expensive it is to keep a group of people in sync.
I build multi-agent systems for a living, and a while ago I noticed I was doing something strange: writing org charts for software. I had phases, gates, checklists and escalation paths. I was giving agents a process copied almost directly from human resources.
Humans developing agent org charts works, but I don't think it is where we end up. My prediction is that we'll move from human-designed development processes to synthetic ones: processes assembled by the agents themselves, for the task in front of them.
Instead of following a workflow we designed in advance, an agent will construct the workflow it needs. It will decide whether a problem should be decomposed, whether an answer needs verification, whether another agent should take a second look, and when a human needs to be involved.
To get there takes a fairly practical form of self-awareness: an agent being able to judge what it is currently good at, and where it is currently likely to fail. Once that becomes reliable, a fixed workflow or org chart starts to make less sense. You can generate a better one for each task, and throw it away once the task is done.
This isn’t about removing humans from the loop. Humans stay exactly where their judgment and accountability matter—and an agent that understands its own limits will call for a human more precisely than any fixed process ever did. It's also not workflow engineering, where we keep drawing cleverer graphs for agents to follow. It's the opposite: The graph stops being something we author and becomes something assembled for the task at hand.
For 70 years we designed processes to compensate for what humans can't do. The next ones will be designed by the agents themselves, around what they can do—today, in this context, on this task.
