Sometime in the middle of 2025, the two smartest agent teams in the industry had a public fight, and both of them won.
In June, Anthropic published a walkthrough of its multi-agent research system: a lead agent that holds the context, spawns a handful of subagents to chase separate threads in parallel, and stitches their findings back together. Their number was blunt. The multi-agent setup, Opus 4 leading Sonnet 4 subagents, outperformed single-agent Opus 4 by 90.2% on their internal research eval, at roughly fifteen times the tokens. Days earlier, Cognition had published the opposite counsel under a title that left no room: Don’t Build Multi-Agents. Split a task across agents that don’t share full context and their decisions quietly conflict, so keep it in one thread.
Read them side by side and the disagreement is narrower than it looks. Both teams already accept that serious AI work is plural. The fight is only about the plumbing.
That is the thing worth noticing. “Team of AI agents” has quietly stopped meaning what you’d think. It no longer describes a staff you assemble and direct. It describes internal wiring: a lead model dispatching disposable workers you will never meet, each one spun up for a narrow job and deleted before you knew it existed. The org got built. You just weren’t invited to look at it.
The summary string and the vanish
Here’s the shape everyone converged on, whichever side of the plumbing debate they landed on. A lead agent owns the context and the plan. It decides how to carve up your request, how many helpers to spawn, which model each one runs, and what counts as done. The helpers do their narrow piece, hand back a paragraph, and cease to exist. You get the final answer.
Notice what you never got. You never saw the carve-up. You never named one of those agents. You never chose which model handled the part of the job that could lose you the account. You never got told any of it happened. You were handed a result and asked to trust a management you weren’t allowed to watch.
That is not a super-assistant. It’s an org chart turned to face the wall, run by a manager who isn’t you.
An org you can’t see is still an org
The reflex is to call this a convenience. One box, one answer, no dropdowns. But a plurality you can’t inspect doesn’t remove the management problem. It just moves the manager offscreen and hopes you won’t ask who they are.
When one of those subagents gets something wrong, and at fifteen times the tokens something will, you can’t find the one that did it. There’s no associate to correct, no seam to reach into. You get an apology from the front desk and a fresh roll of the same dice. A director who can’t see their staff or reassign the job that went sideways is not directing anything. They’re a spectator to an operation held in their name.
OpenAI’s framing of a single personal super-assistant is the most honest version of this, precisely because it drops the pretense of a team. One face, everything behind it. Tidy. Also a promise that the org will stay invisible on purpose.
Whose org chart is it
Here’s the disagreeable part, and it’s a real fork, not a both-sides shrug. The industry treats the invisibility as the feature: fewer decisions for you, less surface to manage. I think the invisibility is the whole loss. The moment the work became plural, you needed to be able to see the plurality, or you stopped being a director and became a suggestion box.
Brief’s bet runs the other way. The associates are standing, not summoned and shredded. They have names. You can see which model each one runs and change it in a sentence when the stakes change. The org chart is yours to look at and reorder, because it’s yours. A lead agent can still coordinate the busywork underneath; it does not get to be the one who decides who’s on staff and quietly keep the roster from you.
The frontier already settled the hard question. AI work takes many hands, each doing a narrow job well. It got the easy question backwards. It decided the one person who should never see the org chart is the one whose work it does.
Brief is opening to a small group at a time. Direct a team instead of operating one more tool.
Request access