Multiplayer AI

For the last six months I have been flogging multiplayer AI. Most of the conversations ended in a few minutes with a human dial tone. Multiplayer AI implies many humans working together. Humans in the workplace? Talking to each other? In the future? Madness.

Now a few things have changed. We quietly gas-lit/steered away from agents being autonomously little people in the computer, and towards being tools for users. Good. We understand that pure model reasoning is horribly inefficient and pretty bad relative to human reasoning + model reasoning working together. Anthropic released a really solid slackbot, lending gravity to the category.

Multiplayer AI matters so much because it opens AI up to the most powerful concepts in software: addressability, accessibility, and human network scale effects. To strip out the jargon: software that allows a lot of people to access the same stuff together can change the world. SaaS, social media, video games, and a lot more all run on this principle.

It’s also a much more technically honest way of understanding LLM interfaces. This is an obscure point but I think it matters. An agent is not a self-contained personality with willpower. It’s an application-level continuity pattern built from model calls and conversation state maintained in a growing document-style workspace object that can be made addressable like any other web page or server-owned resource. Wrap your head around that and so many of the seemingly unsolvable questions about agent permissions, etc. melt away. The answer was in front of us this whole time. It’s a computer.

Now we are all in a race to figure out what the future looks like. Whatever shape it takes, it’s not going to be “Slack + AI” or “chat threads with N+1 users.” Those are good first steps, transitional experiments that we can learn from. But they are not the end state.

Why am I confident about this? Partly because I am building and experimenting with these primitives and they are woefully inadequate. None pass the crucial test of showing them to a civilian and sparking recognition. The winning paradigm is not going to be a dense, feature-rich dashboard festooned with controls and labels and fields. It’s going to be something that feels like a few people talking.

What’s the cherry here? What’s the prize? It’s not more workslop, made faster in Slack. That’s an unimaginative view of the future, where AI is just a bot that automates the maintenance of bureaucratic artifacts that never really needed to exist in the first place.

That’s the basic error that Claude Tag makes. Their mindset is firmly anchored in using AI to drive the SaaS apps of yesteryear. Fill out the report, update the dashboard, write the wiki page, tag the conversation, produce one more artifact that mostly exists because we spent decades building software that needed people to translate their own work into software-shaped objects.

The winning version, the version I want to build and use, lets people talk, ask, answer, argue, decide, complain, exclaim, and teach in the ways humans already do. Let AI sort that mess out and turn it into something useful their colleagues can inspect, correct, and reuse.

The output of collaboration should not be more artifacts. It should be memory.

We’re so early. Lfgoooooo.