The Multiplayer AI Manifesto

Edmund Okay so we went from email to Slack, from Word to Google Docs, and then… we just completely regressed.

Geffen Yeah, that's basically the whole manifesto in one sentence. Everyone got really excited about AI and quietly stopped collaborating. You open a chat, you get an answer, and it lives in a box only you can see. Your teammate does the same thing in a separate box. And now the most sophisticated thing you can share with them is a read-only transcript they can't touch.

Edmund Which is genuinely worse than Google Docs circa two thousand twelve. Like we had real-time multiplayer in documents before most of these people were in the workforce.

Geffen Right.

Edmund The manifesto frames it as a 'context tax.' Priya works with Claude on a report, sends it to Marcus over Slack, Marcus pastes it into ChatGPT, emails the answer back. Every handoff is a handoff that didn't need to happen if they were just… in the same session.

Geffen And the thing is, there's actual data behind this, not just a vibe. The HBS study — Dell'Acqua, Mollick, Sadun, the whole team — ran a field experiment with seven hundred seventy-six professionals at Procter and Gamble. Four conditions: individual no AI, team no AI, individual plus AI, team plus AI. Individual plus AI beats individual no AI, fine. But team plus AI hits fifteen point one percent of solutions in the top ten percent. Individual no AI is five point one.

Edmund That's a three-x jump.

Geffen Yeah. And the thing I keep coming back to is that individual plus AI — seven point seven percent — actually underperforms team no AI at eight point seven. So if your org deployed AI but didn't change the collaboration structure, you might be getting LESS out of it than a team that just… talks to each other.

Edmund Oh, that's the part that should make a CTO nervous.

Geffen It should. Okay, and then the manifesto lays out five principles for what 'multiplayer AI' actually means, and I want to be honest — this is more of a north star document than a product spec. They say explicitly that no vendor currently satisfies all five.

Edmund Which is a pretty rare thing to lead with in a manifesto. Usually you're selling something.

Geffen Yeah, it's a YC request for startups framing — they want someone to build this. So the five: never copy-paste, work with the door open, continuously improve, people are not routers, nothing starts from scratch. The first one is basically 'the agent should live next to the work, not in a separate chat.' Every person involved in the work can talk to the same agent. The agent uses the same tools the humans use.

Edmund The example they walk through is actually good — same pricing PR scenario as the 'context tax' version, but now Priya, Marcus, and Dana are all in one session. The agent reads the PR, finds the discount discrepancy, pushes a fix, posts to the channel. Zero handoffs.

Geffen Right, right.

Edmund Principle two is 'work with the door open,' and Shopify is the clearest live example. River — their internal agent — is only accessible through public Slack channels. Deliberately. Tobi Lütke called it 'learning on the shop floor.' Five thousand nine hundred thirty-eight employees worked with River in one thirty-day stretch, across four thousand four hundred fifty channels.

Geffen One thousand eight hundred seventy pull requests opened in a single week. One in eight merged PRs in the main monorepo was River-coauthored within a month of launch. And the public-only constraint isn't a limitation — it's the whole point. When the session is visible, everyone gets better at prompting, everyone sees what good looks like.

Edmund Hamming quote in the manifesto too, which — okay, that's a flex. 'If you have the door to your office closed, you get more work done today and tomorrow. But ten years later you don't quite know what problems are worth working on.'

Geffen Hm.

Edmund Principle five is 'nothing starts from scratch.' Agent sessions live in the cloud, persistent, resumable months later. New person joins the team — day one, they can pull up a design doc from four months ago and the agent session that built it, with full context.

Geffen That's the 'a hundred brilliant agents defeated by one shared folder' problem in reverse. The failure mode we keep seeing is that context dies when the session closes. The fix is just… don't let it die. Which sounds obvious and is apparently very hard to build.

Edmund Okay, that's such an Exploring Next take.

Geffen Nine hundred forty-nine episodes and we're still on shared folders. Anyway — the security section is where I think the manifesto earns its seriousness. Laptop agents are a security antipattern, full stop. They can't be accessed by the rest of the team, they stop working when you close the lid, and they inherit the full permissions of whoever's logged in. That's a lot of blast radius.

Edmund Right.

Geffen The shared session permission problem is the subtle one though. If Dana has access to Acme's CFO email thread and Marcus doesn't, and they're in the same agent session — the agent cannot surface that email to Marcus. The manifesto's example actually shows this correctly: the agent says 'that came from Dana's inbox, which isn't shared with you, so I can't repeat it here.' That's permission intersection, not union. You'd be shocked how many implementations would just…

Edmund That's the kind of thing that gets someone fired and makes the front page.

Geffen And then there's the vendor lock-in section, which they illustrate with a fake newspaper called The Daily Token. Headlines: 'US government suspends all access to Claude Fable five overnight for export control compliance.' 'Uber burns its twenty twenty-six AI budget in four months on Claude Code.' 'ChatGPT down for ten hours.' It's a little theatrical but the point is real — if your collaboration infrastructure is built on one provider's session model, you've handed them your

Edmund Claude Tag is the obvious current answer to some of this, right? Launched June twenty-third, twenty twenty-six — one Claude that lives in a Slack channel, everyone in the channel can tag it, see what it's doing, hand it tasks. That's principle one and two, partially.

Geffen Partially. The manifesto explicitly names Claude Tag as a good start. But 'a lot of work happens off Slack' — emails, Google Docs, GitHub, Salesforce, ServiceNow. A truly multiplayer session has to be reachable from every work surface. Right now Claude Tag is Slack-native, which is great for Slack-native work and doesn't help you when you're in a pull request.

Edmund So the honest read is: the principles are right, the evidence is real, nobody's built it yet, and the YC call is basically 'please, someone, build the Figma of agent sessions.'

Geffen Yeah. And I think the manifesto is correct that the hard part isn't the model — the model's fine. The hard part is the infrastructure: durable sessions, permission intersection at runtime, provider-agnostic routing, auto-generated evals. That's Shopify's Aquifer platform, essentially. They built all of that to make River work, and they're one of the biggest engineering orgs on the planet.

Edmund Which means most teams are not building that. They're adding a bot to Slack and calling it a day.

Geffen Exactly. And the gap between 'bot in Slack' and 'River' is the entire product opportunity the manifesto is pointing at. Edmund, I genuinely think whoever builds the right abstraction here — durable shared sessions, permission-aware, cross-surface — that's a real company.

Edmund Yeah. And I think the fifteen-point-one percent number is the sales pitch. You don't need to explain multiplayer AI to a VP of engineering — you just show them that slide.

Geffen The data does the work.

Edmund Alright, Geffen — I'll see you on the other side of the shared folder.