Multiplayer AI: The Loneliest Tool in the Office — What It Takes to Put a Whole Team Around One Agent

This is the fifth paper in this series reading Y Combinator's Fall 2026 Requests for Startups one request at a time. The first four moved outward — education, American defense, frontier compute infrastructure, and a cloud for small software. This one turns inward, to the room where the work actually happens. Aaron Epstein's request, "Multiplayer AI," is unusual for the batch because it asks for no new capability at all. It asks for a missing chair.

What the request actually says

Epstein opens with a history lesson: "The best work tools of the last two decades won by going multiplayer. Google Docs replaced Microsoft Word. Figma beat Photoshop." Then the turn: "AI agents are the most powerful new tool a team has, but it's the one thing people still use by themselves." In his telling, the current shape of agent usage — open a private chat, get an answer, paste a read-only transcript to colleagues — is the Microsoft Word stage of the technology. The work product moves between people only as a dead artifact; the live session belongs to one person.

What he wants instead is concrete: shared agent sessions that a whole team can drop into, watch mid-task, redirect, and hand off the way human collaborators hand things off — multi-day agent work with collaboration built in rather than scattered across "a thousand private threads." The reasoning for timing is a single sentence: as agents take on work spanning hours, days, and weeks, "work at that scale was never meant to be done alone, and pulls in many people." The wedge advice is equally compact — "anywhere a team already crowds around one problem, there should be multiplayer agents" — and the request closes with a call for founders "building AI that's multiplayer by default."

Why this request, and why now

The argument holds together because the chat box was correct for its moment. When a model's useful output took thirty seconds, a private window was the right container: one question, one answer, one person. But the frontier of agent work has shifted from answers to tasks, and tasks of any real length are organizational objects. They accumulate context, require judgment calls, cross role boundaries, and need review before anything irreversible happens. A single-player session at that scale is not just awkward — it is a silo. The knowledge of what the agent was told, what it tried, and why it turned lives in one employee's transcript, and it leaves the company when they close the tab.

There is also a colder market logic under the Docs and Figma analogy, and Epstein surely knows it. Multiplayer was never merely a feature; it was the distribution mechanism. A shared document recruits its own next users — every share link is an acquisition event, and procurement flips from an individual's preference to a team's default. Whoever makes agent sessions genuinely shared inherits that flywheel. Investors beyond YC have converged on the same framing; Andreessen Horowitz's Big Ideas list for 2026 reportedly declared that this is the year AI "unlocks multiplayer mode."

The analogy deserves one honest caveat, though, because it hides the entire difficulty of the problem: Docs and Figma synchronized human edits to a passive artifact. Nobody ever had to ask what the document wanted to do next.

What is actually hard

The third participant in a multiplayer agent session is autonomous, and that breaks most of the collaborative-software toolkit. Conflict-free replicated data types can merge two humans typing into the same paragraph because text does not have intentions. An agent mid-task is executing a plan. When two teammates redirect it simultaneously in incompatible directions, that is not a merge conflict — it is a management problem, and the system needs an answer that software has never really needed before: instruction arbitration. Whose steer wins, under what role, and how is the loser's input preserved rather than silently dropped?

Authority is the second wall. Watching an agent is cheap to build; redirecting and approving are permissions problems. In a single-player session, the owner of the chat is trivially the authority for everything the agent does. In a shared session, "can see it" and "can point it somewhere" and "can approve it spending money" are three different rights held by different people, and they must be enforced per action, not per document — because some of those actions, once taken, cannot be reverted. An edit can be undone; a sent email cannot.

Then there are interruption semantics. "Redirect mid-task" sounds like a feature checkbox until you ask what happens to the work in flight. Does the session branch, roll back, or absorb the correction and continue? This is version control for actions rather than text, and it interacts badly with irreversibility: the checkpoint you would love to rewind to may sit on the far side of something that already touched the outside world. Provenance follows directly. A multi-day, multi-person session is only handoff-able — and only defensible when something goes wrong — if the transcript records durably who told the agent what, and which human authorized which consequence.

Finally, the presence model itself is unsolved. Figma-style live cursors are the wrong default for work measured in days; nobody watches a three-day task in real time. The honest shape of the product is mostly asynchronous with occasional live moments — closer to a pull-request thread crossed with an operations console than to a shared canvas — and getting the notification granularity right, so five stakeholders stay oriented without being buried, is a design problem at least as hard as the sync engine.

Who is attempting it, and what building it takes

The label already has claimants. Dust, the Paris-based enterprise AI platform, has reportedly adopted "multiplayer AI" as its literal positioning — humans and agents collaborating as co-contributors in a shared team environment rather than in individual chat sessions. The coding-agent companies are drifting the same way from another direction, moving agents out of private IDE panes into shared channels where a team can watch and steer runs. And the model labs loom over all of it, since a shared-session feature is an obvious addition to any first-party chat product. The Docs analogy cuts both ways here: it was Google, an incumbent, that shipped multiplayer documents — but it was Figma, a startup, that beat Photoshop. The defensible ground is unlikely to be the sync layer, which will commoditize. It is more likely the workflow wrapped around the session: the permission model, the audit trail, and depth in one specific ritual where, in Epstein's phrase, a team already crowds around one problem — incident response, a deal desk, a product launch.

Concretely, building this takes an event-sourced durable transcript as the source of truth, a synchronization layer on top of it, per-action authority rather than per-session ownership, sane branch-and-interrupt semantics, and a sales motion that starts with a team rather than a seat. None of it requires a better model. All of it requires product decisions the chat box let everyone defer.

Where Gwen stands

This request sits adjacent to Gwen, and it is worth being precise about which half is real today. Gwen's workspace already runs work the way Epstein assumes work will run: as long-lived missions rather than disappearing chats, with durable transcripts that survive reloads, progress a customer can watch and redirect mid-course, and human approvals gating any outward action — sending, publishing, spending. The session is already a place with memory and per-action authority, which is the piece of the multiplayer problem that matters most when something irreversible is on the line. What Gwen has not built is the other half: true multi-user simultaneous sessions, several people inside the same mission at once, each with their own identity and standing to steer. Today a mission is watched and steered by the customer, not by a crowd. The honest read is that the request's direction matches where workspaces like Gwen's are heading — persistent, watchable, permissioned agent work naturally grows a second chair — but growing it is real engineering, not a toggle, and Gwen will not claim multiplayer before it exists.