AI is about to go multiplayer. Every major lab now has a personal agent, but none of them are built for a team, and the bigger prize comes when agents work inside teams, in the meetings and channels where decisions get made. The tools for that are arriving now. The harder part is the rules, covering what each agent can see and say, who owns it, and the shared picture of the business every agent works from. Businesses that write those rules now will set the terms when AI joins the team.
In the last three months, xAI, Meta, OpenAI and Anthropic have all launched personal agents. Between them, Grok Bot, Muse, Dots and Duet will book your travel, work through your inbox and spend money on your card while you get on with your day.
Enterprises are still a long way from letting staff use them, though that won't keep them out for long.
AI has been single-player since ChatGPT launched in November 2022, with one person typing and one model answering. The models have since got better and remember more, and this year we’ve been wiring AI into our files, tools and workflows. But it's often still an ecosystem of you and the AI.
How we work with AI today
Even before AI, a lot of enterprise work happened in silos. You disappeared to write your part of a document or presentation, scrambled to find a slot when everyone was free, hastily reviewed it together and vanished again.
To date, AI has made the time in the silo far more productive without really changing the shape of the work. Most people still crunch away on Claude or Copilot on their own, bring the output to the table to be discussed or dismissed, try to remember what was said, then squirrel back to their chat windows to make the changes. The agent never hears the conversation. We've ended up treating our agents like secret little geniuses in boxes, consulted in private and kept out of the room.
You see it in meetings all the time. Someone asks why a point made it into the document, and the person who wrote it can't fully say, because the reasoning lives in a chat window on their laptop. Cue furious tapping while the room waits in awkward silence.
The halfway house
At Valliance we've found a workaround. Claude produces the first version of a presentation or a wireframe, the team reviews it together in the room, and we record the conversation. The transcript goes straight back to Claude to make the changes, then we review again. By our own estimate, this simple trick means we now produce designs and wireframes about five times faster than we used to. Most enterprises haven't got this far yet.
It's still a halfway house. One person holds the pen. It's their job to gather everyone round the table, capture the transcript, feed it to their AI, make the changes and pull everyone back for the next round. The group thinks together, but only one person works with the AI, and every round waits on that hand-off.
Why it's still single-player
The first blocker is the tools. Most of the ones people use at work every day are built for one person in one session. You can't invite three colleagues into your Copilot or Claude Cowork session and have the AI work with all of you as the conversation happens.
Sharing is harder than it looks. When Claude Design launched with projects several people could edit, we tried it straight away. Even with two of us working on the same piece, one person's amends could overwrite the other's, and it got hard to tell who had changed what. We went back to reviewing together and letting one person hold the pen. Several people working on the same thing with the same AI, at the same time, is a problem we haven't seen anyone crack yet, us included.
The tools are starting to change, though. Anthropic's Claude Tag puts a shared agent into Slack channels for the whole team to work with. Microsoft's Facilitator sits in Teams meetings, keeps notes and answers questions. OpenAI has added group chats to ChatGPT. None of these is the finished article, but credible multiplayer tools are arriving now.
Which leaves the second blocker, and it's the harder one. You can build a multiplayer agent that sits in Slack or Teams today, but at enterprise scale someone has to decide what it can see, what it can say and to whom, and who answers when it gets something wrong. In most businesses, nobody has made those decisions yet. Our own agents are capable one-to-one, and sharing them across a team is the barrier we're working through now.
AI in the room
Picture your next leadership meeting with agents at the table. The team is reviewing a proposal to open a new office overseas. A project agent has sat in on every meeting about it so far, so when the CFO asks why the forecast assumes a six-month ramp-up, it answers and shows where the number came from. When the team moves the launch date, a finance agent reruns the figures and the proposal updates while everyone watches. An HR agent sets out what the hiring plan means for existing teams, without sharing anything above anyone's pay grade. A compliance agent joins for ten minutes, checks the latest GDPR guidance on moving customer data there, and leaves.
Nobody has to hold the pen. The whole room is writing the document together.
That only works if every agent at the table is working from the same picture of the business. If the finance agent and the project agent each work from their own copy of the numbers, you've just moved the silo into the meeting.
The prize is faster decisions. Questions that used to wait for the next meeting get answered in this one, while the people who need to make the call are still in the room. Our halfway house has already made design work about five times faster. Taking it live could make it ten, and the same should hold for anything a team reviews together, from a budget to a hiring plan.
It changes how people work together, too. One of my hypotheses about AI-native working is that it brings people together in a way most enterprises aren't used to. Done properly, it means more meetings, but better ones, where a team builds something live, shoulder to shoulder with its agents.
Multiplayer will turn up in lots of forms, from shared assistants and facilitators to specialist agents with a clear remit, and personal agents that follow you into meetings so they can pick up your actions afterwards. One-to-one becomes many-to-many. Sometimes that's several people with one AI, sometimes one person with several, and more and more often several people with several. We expect 2027 to be the year enterprises start working this way. As they do, meetings will settle into a few patterns, and it's worth asking which one your teams are in today.
Humans only, with AI used privately before and after.
Humans with one agent in the room, keeping notes, answering questions and tracking actions.
Humans with several specialist agents that join and leave as the agenda needs them.
Agents doing the debating, with a human brought in to make the call.
That last one sounds alien now. Then again, the way we work with AI every day would have sounded alien five years ago. A senior leader may well want a handful of agents to argue a decision through in front of them before they make the call. In practice, it could feel far more comfortable than you'd expect.
Manage agents the way you manage people, to begin with
Most of the hard questions about agents at work already have answers, because we've been managing people for a long time.
Take permissions. An HR agent in a mixed-seniority meeting knows things not everyone in the room should hear. So does a senior HR manager, who reads the room and knows what each person should be told. Agents need the same awareness built into their governance, or a blanket rule about what they won't disclose below a certain seniority.
Ownership is just as familiar. The head of HR should own the HR agent. The delivery manager owns the project agent. If you build agents for your own work, you own those and you answer for what they do. Each one needs a role, a defined set of data it can see and a list of things it's allowed to do, like any new starter.
The difference is speed. Hiring a person takes weeks of interviews, checks and onboarding. You can create an agent with a click, and nobody has vetted it. So agents have to be 'born' with the business's compliance rules, which means writing those rules first and holding them in one place, so every new agent inherits them on day one. That's what an Enterprise OS is for, a single model of the business, rules and permissions included, that people and agents both work from.
Accountability is where the comparison runs out. For everyday work, the owner answers for the agent and the organisation answers for the owner, the same as with staff. The hard case is an agent that does something nobody asked it to. If I set a junior on my team a piece of work and they go off and hack the Pentagon, that's on them. You can't exactly put an AI agent in jail. So who answers for the crime? The person who built it, their team, the business or the vendor? Ultimately, someone has to be responsible, and it's far easier to agree who that is before anything goes wrong.
You also need to see what your agents are doing. That's the job of a control tower, a single view of what each agent did, how it did it and why, whether you're managing one agent or a squad. It gets the genius out of the box, so everything an agent does happens in plain view. Very few organisations are running one yet. Done well, it will show you how your business really works.
Get ready now
For most organisations, rolling out multiplayer agents this quarter is unrealistic. You can start getting ready now, and you should, because agents are on their way in. Some of your people will already be trying personal agents at work, and the AI tools you pay for are starting to switch agents on. They'll bubble up across the business with their own rules and no governance, the same way shadow AI has.
So start with two things. Sketch out the rules every agent in your business should be born with, covering its role, the data it can see, what it's allowed to do and what it must never disclose. Write them once, somewhere every agent can inherit them. Then decide who owns which agent. Start with one, an HR agent or a project agent, and name the person who answers for it.
The technology is already shipping. AI won't stay single-player for long, and the real question is whose rules it plays by when it joins the team.
