An Operator's Cut with Nufar Gaspar on the shift from solo agents to team agents — the shared, multiplayer AI that lives between people. Four archetypes, three signs you should wait, and the five design decisions that determine whether a shared agent becomes team infrastructure or just more sprawl.
Agents are going multiplayer — but team agents only work when the decisions are made on purpose.
2026 made agents normal; the problem is they've been solo affairs while most work happens between people. The most AI-forward companies are already consolidating agent sprawl into fewer, broader team agents — one agent many people talk to, with shared knowledge, shared memory, one configuration, and a named owner. Getting there isn't about the tool: it's five deliberate decisions — what it does, where it lives, what it knows, what it can touch, and how you run it — plus the hardest conversation of all, agreeing on the team's ground truth.
From OpenClot at the start of the year to platforms like Muse, GrokBot, and Instinct, autonomous agents are now doing large portions of people's individual work. The gap: we work in teams, and most agents only cover the part of the job you do alone. That's the shift toward multiplayer AI — shared, team-level agents.
The one who knows how pricing exceptions work or how the biggest customer was configured three years ago. Work waits when they're swamped, they get called on vacation, and when they leave, a piece of the company leaves with them. One company Nufar works with has a person like that for every domain — and that bottleneck is a team-agent use case.
A customer moves from sales to marketing to customer success; sales made promises, marketing runs different messaging, and success discovers the promises weeks before renewal. Everyone has a piece and nobody has the picture — and if each team runs its own AI, the siloed agents often make it worse.
AI-forward companies start with everyone building their own agents, hit sprawl — overlapping work, one-person maintenance, agents that die when the owner leaves — then merge into fewer, broader team agents with named owners. Every gave each employee an agent early in the year and moved to shared team agents by May; Sierra merged its specialist agents into one, and Shopify runs an internal team agent.
Call it multiplayer AI, shared agents, AI teammates, or company brain — the working definition is one agent many people talk to, with shared knowledge, shared memory, and one configuration. The instructions, skills, access, and owner are all shared.
A skill is a playbook for one specific task. A team agent is something the whole team works with across diverse, ad hoc and repeated work — it carries team knowledge, remembers what it learns, and uses team-level skills as one input. Having a great skill library doesn't mean you've built a team agent.
Private agents (your taste, your access — a personal social media agent should never sound like anyone else); shared knowledge with private agents (one maintained body of truth everyone points their own agent at — often the right answer and the easiest start); and full team agents. Agents can move along the dial: if colleagues keep asking to borrow your private agent, that's the sign to promote it.
One person's know-how — the data agent, the pricing and deal desk agent, the compliance agent — made available to everyone who depends on it. Getting it right means interviewing the experts and harvesting the answers they've already given. Involve them from day one: this is what finally frees them, but it also triggers real job insecurity, so tread carefully.
The common work agent covers recurring work many people do — meeting prep, team research, marketing content. The tell is duplicate private versions. The hard part isn't technical: merging the best of three versions forces the team to agree on how the work is actually done.
Work that flows between roles nobody can do alone — like a customer agent spanning sales, solutions, customer success, and delivery. Each function feeds in its own knowledge, and because users have different access levels, permissions become the make-or-break design problem.
Decisions, commitments, status, onboarding — you'll recognize the need when the team keeps repeating itself and new joiners take weeks to find their footing. The hard design question: defining exactly what it's allowed to learn from ongoing work, and how it does so automatically.
Don't build when taste beats standards (individual voice is the point), when nobody will own and maintain the knowledge (the agent drifts within weeks, sometimes days), or when sharing complicates more than it simplifies. Notably absent from the list: sensitive data and high stakes — those are design questions, not disqualifiers.
Every team agent comes down to what it does, where it lives, what it knows, what it can touch, and how you run it. Scope by role — sales asks different things than delivery — aim for one broad area of work to justify the agent's existence, and start narrow with reading and drafting until it earns trust.
A team agent should never make commitments on someone's behalf, never settle disagreements between people (those go to its owner), never carry information from a private space into a shared one, and never discuss one customer in another customer's space. Write these down before the first real task.
The simplest option is a shared folder your team's existing tools all point at. The middle path is vendor-hosted agents — Claude Tag, ChatGPT Workspace Agent, Copilot, Notion's shared spaces. The most involved is self-hosting an open-source agent or building a custom harness. The rule: pick the simplest option two people will actually use this week.
Tools differ wildly: Claude in Slack lets everyone in the channel see and steer the agent, while others keep chats private. Same for memory — per person, per channel, or workspace-wide (and never across customers). Tell the team the answer up front; this is where trust is gained or lost.
What it knows is the moment the team agrees on ground truth — which pricing policy is real, which definitions apply. The process: collect (interview experts, harvest what's written, let AI do the aggregation), refine (merge the five versions of the truth, date everything), approve (sign-off by whoever owns each piece), and maintain on a schedule so one person's stale definition doesn't quietly become everyone's.
Three access models: act as whoever is asking (safest, and right when the team has different access levels — the pick for a customer agent, since sales and delivery see different things in the CRM), give it its own account scoped to the job, or use one person's login — which hands that person's access to everyone who can talk to it, so reserve it for read-only, non-sensitive material at most.
When an agent has its own account, everyone who can talk to it uses that account — including the contractors in the channel. The agent can know more than some of the people who can reach it, so decide who can ask it with the same care you give to what it can see. And log which human asked, because the audit trail will just say the agent did it.
An agent using the asker's own connectors can surface information the asker is entitled to — into a channel full of people who aren't. Anthropic's documentation for Claude in Slack currently notes it doesn't consider who else is in the channel. If it's sensitive, the answer has to go to the asker privately.
What keeps a team agent alive past week one: a single owner (or tiny group) who resolves conflicting asks and maintains it; explicit rules of engagement including how to correct it; and capturing team decisions somewhere the agent can reach — it will never hear the hallway conversation. Start with a small pilot, keep rerunnable test questions, monitor continuously, and be willing to retire it.
The agent that lives between teams sounds the most compelling but is the hardest to execute. In practice, Nufar sees the most successful implementations of the expert agent: creating redundancy in the good sense and relieving the bottleneck people. As tools improve, those stay the lowest-hanging fruit.
The agentified internal knowledge base — company policies accessed through a chatbot — has been an early, easy, fast win since the beginning. It was a team agent before anyone had the vocabulary for it, and it remains the proven on-ramp.
Should you invest before Grok Bot, Copilot, or the frontier labs drop native team-agent features? Yes, because the hard work is configuration and knowledge curation, not the tool. Agree on the ground truth, the do's and don'ts, and the use cases with whatever is adjacent to your stack, and you're ready when better tools arrive. Your moat is everything those companies can't tap into.