Your AI is single-player. That's the whole problem.

Every tool that won the last twenty years did it by going multiplayer. AI hasn't had that moment yet, and the reason isn't the model. It's the interface.

I want to start with an observation that isn’t about AI at all.

Microsoft Word was a better word processor than Google Docs for at least a decade. Better typography, better tables, better everything a single writer touches. Google Docs won anyway. Photoshop was, and arguably still is, a more capable pixel editor than Figma. Figma won design teams anyway.

Neither of those was a feature fight. Both were interface fights, and both were won by the same move: turning a thing one person did alone into a place several people did it together.

The winner wasn’t the better tool. It was the tool where the work stopped belonging to one person.

Where AI is stuck right now

Open any AI product built for work today and look at the shape of it. There is a text box. You type into it. Something comes back. It is addressed to you, sits in your history, and is invisible to everyone else in your company.

If a colleague needs what you just made, your options are: send them a screenshot, paste the output into Slack, or send them a read-only link to a transcript they cannot touch. That is not collaboration. That is showing someone a photograph of your work.

Meanwhile the actual work is getting longer. Agents now run for hours. They chase forty vendors, prepare a board pack, reconcile a month of payments. Work at that duration was never a solo activity in any company I have worked at. It pulls in the person who owns the data, the person who owns the decision, and the person who has to live with the result.

What we saw in the field

We spent months sitting with event agencies, schools and credit funds: businesses where coordination is the job. The same thing kept happening.

Someone smart figures out a genuinely useful prompt. A coordinator works out how to get the AI to file receipts properly. An analyst works out how to get it to read a covenant certificate. It works. It saves them four hours a week.

And then it dies with them. It lives in their history. When they go on leave, it goes on leave. When they quit, it quits. The company learned nothing. The next person starts from an empty text box.

This is the part that made us rebuild Kryttle. The bottleneck was never model capability. It was that nothing anyone built with AI ever became company property.

The specific thing we changed

In Kryttle, when you build an agent, it is not yours in the way a chat thread is yours. It is an object in a shared workspace with members, a history and an owner, and critically, anyone on the team can add to it.

Here is the moment that convinced us this was the right shape. It happened at an event agency during client operations.

• Nisha, who runs events, asked the team to keep a proper vendor database.
• Meera, a coordinator, built an agent that reads vendor mail and keeps the sheet current. She shared it.
• Rohan, another coordinator, looked at Meera’s agent and thought: this should also fetch each vendor’s rating. He did not build his own agent. He added a job to hers.
• Within a week, finance had built a spend report on top of the same data, and ops had added a KYC step.

Four people, one system, no meetings about it. Nobody sent a screenshot. Nobody copy-pasted a prompt. The vendor database became a thing the company owns rather than a thing one coordinator maintains.

Why this is an interface problem, not a model problem

You could run this exact story on any frontier model. The models are good enough. They have been for a while. What was missing was somewhere for the work to live that wasn’t a private tab.

Three things had to be true for Rohan to extend Meera’s agent instead of rebuilding it:

• He had to be able to see it. Not a transcript. The actual working thing, running, with its jobs listed.
• He had to be able to change it without asking her, without breaking what she built, and without holding her credentials.
• The change had to be hers too. When his rating job runs, she sees it. It’s one agent, not a fork.

Every one of those is an interface decision. None of them is a model capability.

What this makes possible

Once the work is shared, some things follow that we did not fully anticipate.

Context compounds. Because everyone builds in the same place, the system’s understanding of the business gets better with every agent, not just the ones you personally made. A new joiner asks how a venue change works and gets the company’s actual answer, from the company’s actual SOP.

Coordination gets cheap. When one system can see the vendor sheet, the job sheet and the client thread, it notices that Neil has been blocked on Priya since Tuesday. Nobody had to run a standup to surface that.

And the ceiling moves. The agencies we work with stopped taking on more events not because they lacked talent, but because coordination was the constraint. Take the constraint out and the same team runs more.

The uncomfortable part

Multiplayer is harder to build than single-player, and it is harder to sell in a demo. A private chatbot demos beautifully in ninety seconds. A shared workspace only makes sense once there are three people in it.

We have made peace with that. The version of this that wins is not the one that demos best to one person. It is the one that a team of fifteen still opens on a Tuesday in month eight, because by then it holds too much of how they work to close.

That’s what we’re building.