Teams of Dots: What Multi-Agent Coordination Looks Like in Practice

Every always-on agent launched in the last eight weeks eventually arrives at the same second act: one agent is useful, but a team of them is where the real leverage is supposed to be. OpenAI's Dots, xAI's Grok Bot, and Meta's Muse all gesture at multi-agent coordination in their September 2026 announcements, but they describe it with very different levels of specificity — and the gap between "can run multiple agents" and "multiple agents coordinate reliably" is exactly where this generation of products is thinnest.

What OpenAI has actually said about teams of dots

OpenAI's own description, per coverage of the September 29 DevDay launch, is that users "can run multiple Dots as a team." That's the extent of the confirmed detail in the sources available for this series — there's no published specification of how dots divide labor, whether they share context automatically, or how conflicts between two dots working on related tasks get resolved. Given that each dot runs on its own isolated cloud computer and browser, separate from the user's machine and from other dots unless explicitly connected, the more interesting engineering question — how information moves between dots on a team, rather than just between a dot and its human — is one OpenAI hasn't answered publicly yet.

That's a meaningfully different starting point from Specialist Dots, OpenAI's pilot program giving agents organizational identities, credentials, and defined responsibilities across functions like procurement, invoice processing, email marketing, customer support, and commercial contracting. Those are coordinated by design, provisioned through the same systems a human employee would be onboarded into — which Specialist Dots: The Early Pilots Nobody's Talking About Yet covers separately. "Teams of dots" as a consumer/prosumer feature and "Specialist Dots" as an enterprise pilot appear to be two different points on the same spectrum, not the same feature described twice.

Grok Bot's version is more concrete — and more exposed

xAI has published considerably more detail about what multi-agent coordination looks like in its own product. Users build a sidebar of named bots, each with a distinct role — one for email, one for LinkedIn outreach, one for negotiating, one for general use — and when several bots occupy a group chat, xAI's own launch material says they "can message each other, pass work and assign ownership on their own," per reporting reviewed for this series. Multi-Bot Coordination: When Your Agents Talk to Each Other covers that mechanic from Grok Bot's side in more depth, but the headline mechanism — bots coordinating inside a shared chat thread rather than through a hidden backend protocol — is more specific than anything OpenAI has disclosed about dots so far.

The catch is architectural, not just a feature gap: all of a user's Grok Bots reportedly share one cloud computer per account — including its files, browser sessions, and app logins. That's a real tradeoff. Coordination is easier when every bot on a team has access to the same environment by default, but it also means the isolation model that makes a single bot relatively safe to grant broad access to doesn't extend cleanly to a team of them. Dots take the opposite approach: each dot gets its own separate cloud computer, which likely makes coordination harder to build well but keeps one dot's blast radius from automatically becoming every dot's blast radius.

Coordination and isolation pull in opposite directions. Share one environment across a team of agents and coordination gets easier while containment gets harder; isolate every agent and you get the reverse. Neither vendor has published a convincing answer to both at once.

Where Meta's Muse sits, and what to actually watch for

Meta's Muse doesn't currently describe a multi-agent "team" feature in its own materials — each user's Muse operates inside its own Secure VM with the Sentinel agent approving or blocking connected-app actions, a single-agent-per-user model at least as of the September 8 launch and the September 29 small-business expansion. That's a notable design choice by omission: Meta built one of the more detailed permission architectures of the three products and, so far, hasn't extended it into multi-agent territory the way OpenAI and xAI have started to.

For anyone actually planning to run several agents in production — a dot per repository, a bot per department — the questions worth asking a vendor directly, rather than assuming from a launch announcement, are: does each agent get isolated compute, or shared; how is context passed between agents, and can a human audit that hand-off; and what happens when two agents' actions conflict. None of the three companies has fully answered all three yet. Dots vs Grok Bot vs Muse lays out the single-agent permission comparison in detail — multi-agent coordination is the natural next chapter of that same comparison, and right now it's the chapter every vendor has left mostly blank.

There's a practical reason this comparison matters beyond curiosity: whichever coordination model a team adopts effectively locks in its blast-radius profile for as long as it's using that product. A company standardizing on Grok Bot's shared-computer model for a five-bot department rollout is accepting, upfront, that a compromised credential or a misbehaving bot can touch every other bot's files and sessions on that account — a tradeoff that might be entirely reasonable for a small team that trusts its own access controls, and much less reasonable for a larger org with stricter separation-of-duties requirements. A team building around Dots' per-agent isolation is accepting the opposite tradeoff: better containment today, in exchange for a coordination layer that's less mature and less specified than Grok Bot's. Neither choice is obviously wrong, but it's a decision worth making deliberately, function by function, rather than discovering after the fact which model a vendor actually shipped underneath the marketing language.

This is also a good moment to separate marketing language from engineering reality across the category as a whole. "Agents that coordinate" is now a checkbox every always-on agent vendor wants to tick, but the underlying mechanisms — shared memory, message-passing, task hand-off, conflict resolution — are genuinely hard distributed-systems problems that don't get solved just because a demo showed two chat bubbles talking to each other in a group thread. The teams that get real value out of multi-agent setups in the next year will likely be the ones that treat "coordination" as an engineering question to interrogate vendor by vendor, not a feature to assume exists just because the word appears in a launch announcement.

Part 15 of 30 in the ChatGPT Dots series. Previously: The Content Dot: Turning One Transcript Into a Week of Posts. Next: Who Gets a Dot First: Pro, Business Premium, and the Eligibility Map. Browse more builder insights or explore AI skills at aiskill.market.