TL;DR
The short version
Almost every agent teams run today serves a single person — yet roughly half of knowledge work is collaborative, and that's the half agents haven't touched.
The move is toward shared agents with team-owned context, visible work, and live hand-offs, evidenced by Anthropic's Claude Tag, Open Claude, and Y Combinator's Fall 2026 request for startups. You pick where one should live with a four-axis test, then ship it in a four-session sprint — a read of The AI Daily Brief.
Built on The AI Daily Brief with Nathaniel Whittemore (7 Sep 2026). Verified workplace stats from Gensler, Microsoft, and Asana; Anthropic's internal-code figure is self-reported.
The gap nobody's building for
Look at how your team actually uses AI. One person has a research agent. Another built a chief-of-staff agent. Someone runs a coding agent that files PRs. Every one of those serves a single person doing single-person work.
That covers about half the job. The rest is collaborative — and the numbers say it's the larger half.
Gensler 2026 Global Workplace Survey · Microsoft Work Trend Index · Asana Anatomy of Work.
So the collaborative half of work is the larger half, and it's the half agents haven't touched. Nathaniel Whittemore's framing on the AI Daily Brief: the shift is from single-player AI to multiplayer AI.
What multiplayer actually means
Multiplayer AI is not a personal agent you share a transcript from. It's one agent that lives in the team's space, with context the team owns.
Whittemore is specific about the context part: one shared repository across the team, not everyone's folders holding copies of the same docs. The other three shifts are just as concrete — private outputs become visible work, feedback-after-the-fact becomes live participation, and personal memory becomes shared memory that belongs to the channel or the project.
The clearest test is Anthropic's Claude Tag: one Claude scoped to a Slack channel that everyone tags, carrying the channel's context and picking up where the last person left off. Open Claude's maintainers hit the same wall building 2.0 — passing transcripts around wasn't enough. Once you can open the same live session, it becomes a shared piece of work another trusted developer can inspect, steer, or take over.
This isn't only a coding pattern. Y Combinator named multiplayer AI a Fall 2026 request for startups, pointing at the obvious history: Google Docs beat Word and Figma beat Photoshop by going multiplayer.
AI agents are the most powerful new tool a team has, but it's the one thing people still use by themselves.
Aaron Epstein, Y Combinator
Where to put your first shared agent
Not every overlapping task is worth an agent. Score each candidate workstream 1 to 5 on four axes, then start with the highest total:
- Shared need — how many people need the same context?
- Staleness cost — how much does it hurt when each person's version drifts?
- Permission sensitivity — how much restricted data does it touch?
- Checkability — can you tell quickly whether the agent got it right?
The mechanics are four sessions: inventory what your team runs today, extract the context worth sharing, map where your work actually overlaps, then ship one shared agent used by at least two people on real work. The last step is a loop — if a use case isn't the right shape for an agent, drop it and score the next one.
The boundary worth naming
Two things temper this. First, permission sensitivity is one line on a scoring sheet, not a solved problem. A shared agent with channel-wide data and tool access is exactly what an IT or security lead will want a real answer on before it touches anything restricted. Decide that up front, not after.
Second, the source. This came from an AI-advocacy show that doubles as a launch for the host's own free program and paid courses, so the "top 1%" language is marketing. But the trend under it doesn't depend on the pitch — Gensler, Microsoft, Anthropic, and YC are all independent of the show, and they point the same way.
The operator's takeaway is small and testable: pick one workstream two people already share, put a single agent on it with shared context, and see if the work gets better. That's the whole experiment.
