Multiplayer AI.

Y Combinator's brief on why AI hasn't had its multiplayer moment yet — and which mental models hold, which break, and what new ones emerge when agents go collaborative by default.

▶ Play video

Video: Y Combinator / YouTube

What this brief is really about.

Every platform shift has a tell: the moment when the solo tool becomes the shared surface. VisiCalc was a solo tool; Excel became the shared spreadsheet. Email was solo; Slack brought it into rooms. Word was a single-author document; Google Docs made editing a sport for a crowd. Y Combinator's two-minute brief on multiplayer AI identifies the precise equivalent inflection point for AI agents — the moment we realise we've been using the most powerful new tool as a private scratchpad when it should be the team's living workspace.

The argument is deceptively simple. AI today is architecturally single-player: one user, one chat box, one context window, one result you can share only as a static read-only link. But tasks at the agentic frontier — the ones that run for hours or days — were never designed to be done by one person alone. They pull in reviewers, decision-makers, hand-off recipients. The coordination machinery that grew up around those long-horizon tasks assumes multiple people can see and touch the work as it unfolds.

Three kinds of model edit are on offer in ninety-four seconds. Classic coordination models come out amplified. Some get quietly reversed. And at least one genuinely new model earns a place in the latticework.

Old models, sharper edges.

The YC brief leads with two of the cleanest illustrations network effects have ever had in a product context. Google Docs beat Word not because it was a better word processor, but because network effects accrued every time a second person opened the same document. Figma beat Photoshop on the same dynamic: design review stopped requiring email attachments when everyone could look at the same canvas simultaneously. The observation that "the best work tools of the last two decades won by going multiplayer" is a crisp restatement of this classic model — and the implication is that the same dynamic is about to replay in AI.

Coordination costs are the second amplified model. The classic formulation says that as teams grow, coordination overhead grows faster than the team. This brief argues the inverse: reduce the visibility barrier between teammates and the agent's working context, and coordination costs collapse. "Anyone on a team should be able to drop into the same live agent session" is not a UX feature; it's a coordination-cost claim.

Leverage gets extended in a specific direction: not just the leverage of an individual using AI, but the organisational leverage of a shared agent session. When the same agent context is readable and steerable by multiple teammates simultaneously, the leverage multiplier applies to the team, not just the user.

Models that do not survive intact.

The model most quietly overturned here is asynchronous handoff as the default mode of knowledge work. The current paradigm: you finish a piece of work, hand a link to the next person, they start from the artifact. This breaks cleanly for tasks that are short and discrete. For long-horizon agentic tasks — days of running, not minutes — the clean handoff never comes. The agent is always mid-stream. "Send a link to a read-only transcript they can't touch" is the brief's description of the current state; the framing makes clear this is a failure mode, not a feature.

Specialisation of AI tools by user also comes under pressure. The implicit current model: each person has their own AI context, tuned by their own prompts, inaccessible to others. The multiplayer frame flips this. Just as no one person "owns" a Google Doc, the suggestion is that the agent's working context should be a shared resource. The mental model of "my AI vs. your AI" gets replaced by "our agent."

And finally, task completion as a clean event bends. The phrase "shared living thing instead of a thousand private threads" reframes completion. Agentic tasks at scale don't end; they hand off. The latticework entry for what constitutes a finished unit of AI work needs updating: a handed-off, steerable, inspectable in-progress session may be more valuable than a delivered artifact.

New entries for the latticework.

The brief's central coinage is multiplayer-native by default — the principle that AI tools should assume collaboration from the first design decision rather than bolting multi-user features on after the fact. Google Docs wasn't "Word with sharing"; it was designed for concurrent editing from the start. The multiplayer-native model predicts that AI tools designed as single-player with sharing features will lose to tools that treat shared context as the unit of state, not an afterthought.

The examples in the final twenty seconds — engineers, sales, support, legal, finance, marketing — gesture toward a new model worth naming: role-ambient agents. The claim is that every function that currently "crowds around one problem" is a candidate for a shared agent. The shared agent is not a product for a specific role; it's infrastructure for any context where a team's attention converges. This is a different claim than "AI for sales" or "AI for legal" — it's a claim about the shape of coordination, not the domain of expertise.

The corollary new model is ambient presence in long-horizon work. Current AI gives you presence in a session: you're there for the duration, then you leave. Multiplayer agents need a different presence model: drop in, observe, redirect, hand off, drop out — without losing continuity. The model to reach for is less "conversation" and more "live document." A shared agent session is closer to a Google Doc that also does things than to a chat thread that delivers answers.

Finally, the coordination inversion: historically, coordination tools followed work tools (Slack added to email, Zoom added to calls). The multiplayer AI thesis is that coordination should be built in from the start — that the session itself is the coordination artifact. This inverts the usual build order.

When to reach for which.

The latticework, after.

The brief is ninety-four seconds and makes one claim with unusual precision: the last two decades of platform shifts had a common mechanism, and AI has not yet triggered it. Whether that claim proves true depends on execution nobody can preview yet. But the latticework update it implies is worth making now, before the fact: when shared context becomes the unit of AI state, coordination models that assumed separate contexts will need re-rating — and the tools that win will be the ones that built multiplayer in, not on top.

This turns the work a team does with agents into a shared living thing instead of a thousand private threads. Y Combinator, Multiplayer AI