What is multiplayer AI?

Why is Anthropic betting on multiplayer AI?

Scott White, head of enterprise product at Anthropic, argues the biggest bottleneck in enterprise AI is not model intelligence but the fact that most people still use AI alone. The August 2026 Claude Tag update shows what that bet looks like in practice. Anthropic removed the classifier that judged each Slack message in isolation. Claude now reads the channel’s full context, plus its memory and standing instructions, and picks one of four moves: reply inline, start deeper work in a thread, route the message into an existing workstream, or say nothing. Anthropic says the change made it roughly 30% better at deciding when, and when not, to interject (VentureBeat, August 2026).

White’s framing of the shift is poignant. “Claude used to feel like your personal chief of staff,” he told VentureBeat. “Now Claude, in the context of organisational deployment, feels like the company’s chief of staff.” The restraint is deliberate: “an annoying agent is worse than an unhelpful one,” and Claude goes dormant in channels where it repeatedly has nothing to add.

What made proactive AI agents possible?

Proactive agents became viable when three things arrived together: MCP giving governed connectivity into enterprise systems, models crossing an intelligence threshold where proactivity stops being annoying, and a form factor that puts the agent where collaboration already happens. White calls MCP “the USB-C for AI connectors”, and it’s the open standard even Anthropic’s rivals have adopted.

Underneath sits an evolution he describes in three phases. AI handled a part of one task, then whole tasks, and now what he calls projects or goals: keep the product bug-free, make NDA review faster. His forecast goes further: “Soon, I think, we’ll give Claude its OKRs, and it will figure out which projects the organisation needs to work on and how to connect people to do them.”

Goals are where multiplayer becomes unavoidable. A task can belong to one person. A goal belongs to a team, its systems, and increasingly its agents.

What’s missing from the multiplayer AI story?

The missing piece is the model side: the coverage frames multiplayer as many humans working with one agent, and says almost nothing about many agents working with each other. Once a goal has more than one agent on it (a Slack agent, a coding agent, a CRM agent, often from different vendors), each of them needs what a new hire needs on day one: the organisation’s context, its rules, and governed access to its systems. Ship that inside one product and only that product’s agent gets it. Every other agent starts from zero.

The adoption numbers say connective tissue is exactly what’s absent. McKinsey’s State of AI survey, cited in the same VentureBeat piece, found 88% of organisations use AI in at least one function and 62% are experimenting with agents, yet only 39% attribute any earnings impact to AI, and just 6% qualify as high performers. Add the pricing uncertainty White alludes to (Claude Tag’s expanded channel context is free “for now”, with unit economics unsettled), and a strategy question follows. In my view it’s the first one to ask: if your organisation’s working context lives inside one vendor’s agent, what do you own when the meter turns on, or the model changes?

How do you make AI multiplayer for the models as well?

You make AI multiplayer for the models by giving every agent, from every vendor, one shared and governed layer: contexts it can load, connections it can act through, workflows that run the same way every time, and guardrails enforced outside any single tool. This is what Oi builds. Connect once over MCP (live connections today include Linear, GitHub, Figma, Xero, Google Analytics, Google Workspace, Neon, and Zoho CRM, credential-brokered so runtimes never hold provider credentials), start from hundreds of public contexts, and every agent that shows up for your goal loads the same brain.

Concretely: a contract administrator on a construction job has one agent watching the project channel and another drafting the response to a variation. Both need the committed cost from the right system of record, the margin floor they can’t quote under, and which numbers are client-safe to send. Carried in a shared layer, both agents give the same answer. Keep it per-tool and you’ve got confident systems acting on partial pictures.

Models will keep changing. Your context shouldn’t have to. The human side of multiplayer AI is shipping now, and the model side is buildable today: connect once, load proven contexts, govern at the layer, and let the work compound.

Book a demo

Frequently asked questions

What is the difference between multiplayer AI and a chatbot?

A chatbot answers one person’s prompts in a private session. Multiplayer AI works across a team: it reads shared context, joins conversations on its own judgment, and pursues goals that involve several people. Anthropic’s Claude Tag in Slack is the best-known example.

What is Anthropic’s Claude Tag?

Claude Tag is Anthropic’s agent that lives inside Slack channels. An August 2026 update let it read whole conversations instead of single messages and choose between replying inline, opening a thread, routing to a workstream, or staying silent. Anthropic reports it is about 30% better at deciding when to join in.

What do multiple AI agents need to work together?

They need shared context about the organisation, governed connections into its systems, repeatable workflows, and guardrails enforced outside any single tool. Without those, each agent acts on its own partial picture of the business.

Does multiplayer AI require one vendor’s ecosystem?

No. The Model Context Protocol (MCP) is an open standard adopted across major vendors, so contexts, connections, workflows, and guardrails can live in a neutral layer the organisation owns and point at any model.