What do you mean by using agents at scale? How far can you go?


When I say "at scale," I mean going well past one agent finishing one task. Not that long ago, a single agent finishing a coding task felt like magic. Now we're orchestrating thousands of them across devices. The most eye-popping recent example: OpenAI pointed 10,000 concurrent agents at a Millennium Prize problem. That's a lot of agents.

The progression looks something like this: code completions, then chat, then agents using tools, then lots of agents using tools at once.

At the far end, you manage agents, they manage agents, and those manage more agents. In Intent, that means coordinating multiple agents for each task, isolating multiple tasks into workspaces, and running workspaces across multiple devices.

So how far can you go? Technically, pretty far, as long as you solve coordination. Agents need to avoid stepping on each other's work, communicate among themselves (say when they're done, delegate, review each other's output), check in with humans when needed, and get through workflows in parallel, in sequence, and asynchronously. We do this with isolated workspaces, focused context, specialized roles (coordinator, implementer, verifier), and handoffs where an agent that's done passes work to a fresh agent that only carries important information forward.

The harder limit is us. Think of an org chart: a manager keeps tabs on a handful of reports, a director manages managers, a CEO manages directors. At each layer you trade direct understanding for leverage. Being a good CEO is rare. A CEO of CEOs? Vanishingly rare. How many layers of agent delegation a person can meaningfully oversee and course-correct is still an open question.

That's also why the interface has to change. Chat threads break at that scale, so you need visual reviews of what agents changed, diagrams of how the system works, and views you can arrange however you want. Put simply, the number of agents you can run is mostly a coordination problem. The number you can actually steer is a design problem.