Where you do work is changing.

At the Dreamforce 2026 Slack keynote, Boris Cherny (the head of Anthropic’s Claude Code) approximated that 90% of his coding is done in Slack. Why is he using a chat tool to code? And how? Forgive the immediate use of a buzzword, but the answer to both questions is “multiplayer AI,” a slightly opaque term that’s simpler than you might expect.

The idea behind multiplayer AI is to bring humans, AI, and “context” together into a persistent single place for everyone to work together. “Context,” in this regard, includes conversation history, previously shared documents, permissions, organizational structure, prior decisions, active projects, etc. For more on context, please see Boris Evelson and Indranil Bandyopadhyay’s work: The Next Evolution Of AI Will Rely On Context Layers.

Admittedly, the name multiplayer AI is imperfect. “Multiplayer AI” can sound like several AI agents operating together. But the defining feature of multiplayer AI is human participation. If my colleagues and I are all working with an AI agent in a shared place (like a chat tool or even a Word doc) where everyone can see what AI — and others — are doing, ask AI or each other for help, and pick up with AI where a colleague left off, that’s multiplayer AI. Multiplayer AI can help facilitate agent-to-agent communications, but the most important part is multiple humans interacting in a shared surface with AI.

How does multiplayer AI work in practice?

In application development, to use Mr. Cherny as an example again, he leverages a Claude/Slack integration (Claude Tag) that allows Claude Code to live in Slack alongside Mr. Cherny and his team. Now, Claude can be dispatched to generate code and come back with an update in a shared chat thread that everyone — Mr. Cherny, other users, or even other agents — can engage with.

And this doesn’t just apply to coding. Multiplayer AI is poised to transform collaborative work, big and small. Instead of hunting through a dozen systems to find a data point, manually updating a client record, or constantly swiveling between a chat tool and creative suite to make 14 last-minute adjustments to an image, users can do all this work from a single place. According to a sales development rep using multiplayer AI, they were spending 90% of their day in Slack and only spending 10% of their day in “single-player” interfaces.

Why should I care about multiplayer AI now?

Beyond “it’s a thing,” why should you be thinking about (and planning for) multiplayer now? Well frankly, because multiplayer AI is going to be a standard way of working in the future. In fact, there’s a good chance that this standardization will happen over the next year.

There are three forces driving this:

1. Work is already very multiplayer today.

We humans spend a lot of time working with others. Employees report spending an average of 12 to 15 hours a week in meetings (a day and a half). This number doesn’t include “async collaboration” channels (email, chat, etc.), and often we’re collaborating with more than one person at the same time. Just think: How often are you on a video call with multiple people? In contrast, AI that’s trapped as an individual assistant cannot see enough of the team’s work to contribute effectively. AI right now is just adding another isolated interface that increases collaboration headache. AI needs broader collaboration capabilities and context to be useful.

By the way, an interesting side note that came out of the Hugging Face attack was that OpenAI’s agents spontaneously created their own message board (by hijacking a package repository). They then used the appropriated file system to coordinate work, share notes and tools, and act as a persistent source of context as agents came and went. This is unexpectedly good proof that distributed sources of context will make agents’ jobs harder and that multiagent systems will need clear coordination channels to actually manage more complex work. And it’s best if humans can also be involved, instead of trusting a Minions-esque swarm of agents to go off and do work without adult supervision.

2. Vendors are making multiplayer AI just kinda built in.

While Slack is eagerly embracing its role as a multiplayer AI shared surface (as evidenced by the recently launched Slack Code, Slackbot, and Salesforce’s headless strategy), it’s not just Anthropic and Slack doing this. Your shared surface of tomorrow could be a team chat tool, a shared document that allows multiple people to simultaneously edit, or some other “canvas” where people collaborate in real time.

Other vendors (in alphabetical order, nonexhaustive) such as Asana, Atlassian (which recently introduced an Agentic Multiplayer Protocol), ClickUp, Glean, Google, Microsoft, OpenAI, ServiceNow, Zoom, and more are championing and building multiplayer AI tools. As these come online, the trend of “it’s already there” from key tools will lead most organizations to accidentally adopt a multiplayer AI approach. Before this happens, I’d recommend testing multiplayer AI postures before it pops up suddenly in an unwelcome space, because everyone loves spontaneously appearing chatbots (RIP, Clippy: You tried your best, and I’ll always appreciate you for that).

3. Our current way of working is horrifyingly inefficient (and organizations proved that multiplayer AI works back in 2013).

Unfortunately, the way we work is bad. Eighty-six percent of information worker decision-makers reported using over four tools a week, and 43% reported having to use more than seven. It’s no wonder that 46% of employees struggle finding critical information every month (or more often). Add collaboration fragmentation across email, chat, meetings, etc., and it’s a miracle we can work at all. And we know this is a problem. We’ve (kinda) already solved it before — back in 2013, with ChatOps, a direct forerunner. Then we did it again in the data space with DataComs (2022) and curiosity velocity (2023).

The ideas behind each iteration are effectively identical to multiplayer AI: Turn shared chat interfaces into work hubs by extending them into multiple tools (systems of record or action) so people can automate and control work from Slack or eventually Teams. The issue was that this wasn’t universally accessible — you needed to know what an API was. With AI agents (and systems that are easier to connect to), this moat has effectively been bridged.

In short, multiplayer AI is needed to make AI useful to people, vendors are making it borderline a default, and this all comes with the definitive benefit that it makes work less of a context-switching nightmare. It’ll get introduced, and once it proves (once again) to make people’s lives better, it’ll stick. So why wouldn’t someone want to work this way? For some of what can go wrong (and how we can start to prepare), please see the forthcoming second part of this blog.

Have questions on multiplayer AI and what it means for your organization? Clients can schedule a guidance session with myself (Will McKeon-White) or Kim Herrington. Leverage Forrester AI for instant insight!

Share