A few days ago, we released the Multiplayer AI Manifesto, sharing our vision for AI at work. Multiplayer AI lets teammates work with the same agents in shared sessions, learn from each other, and control what data agents can access. Today, most AI tools leave each person working alone in a private session.
Why is Multiplayer AI important?
In recent years we have made an increasing number of our work tools multiplayer: Google Docs, Notion, Figma, and many others. Multiple people can edit, comment, and see the same work.
But recently, we moved much of the work we do into siloed AI chats. Sharing what you figure out with an agent takes extra work.
Someone copies an answer from Claude into Slack. A teammate pastes it into their own agent, asks a follow-up question, and emails the answer back. Every handoff means explaining the task again.
And when AI sessions are private, learning is slow. If you only learn from the prompts and workflows you develop yourself, you learn at one person's pace. But if teammates could see each other's questions, corrections, and results, everyone would get better faster. We are all still figuring out how to work with agents. We should do it together.
Security is another reason to make this shift. Agents can take actions, and an agent on an employee's laptop may have access to sensitive files and credentials. Running agents in shared cloud infrastructure gives the organization a place to limit data access, firewall network connections, and track what they do.
What makes AI truly multiplayer?
Teammates participate in the same session
Shared billing, credentials, and skills are useful, but the real test is whether a teammate can talk to the agent you are already working with. Can they ask why it made a decision? Can they add a requirement and ask it to make a change? Sharing a transcript is only the beginning.
You can reach the agent where the work happens
Adding an agent to Slack is a good start. But a lot of work happens outside Slack. When you are coding, you want the conversation next to the code and the live app. When you are writing a document, you want the agent next to the document. The goal is to reach the same session from every place you work.
Teams share what they learn
At a minimum, people should be able to learn from each other by seeing which techniques work best. Ideally, agents should also improve automatically as they work with the team. An agent launched tomorrow should benefit from what the team learned today.
The five principles of Multiplayer AI
We wrote the Multiplayer AI Manifesto to describe a vision for what this experience should become. Its five principles guide what we're building. No platform, including Superconductor, fulfills all of them today.
1. Never copy-and-paste
Agents must live next to the work. Everyone involved should be able to talk directly to the same agent, and the agent should be able to use the tools the team uses to get the job done.
2. Work with the door open
When a teammate writes a great prompt, everyone should be able to learn from it. Keep the process visible: the questions, the wrong turns, and the insights. We will all learn faster with our doors open.
3. Continuously improve
When an agent needs several rounds of correction, that learning should automatically become a reusable skill. Recurring work should have an eval or benchmark, so the team can see whether its skills and agents are actually getting better.
4. People are not routers
A human should never be asked a question that an agent already has the answer to. Chasing project updates and relaying answers is work for agents. People should focus on the decisions that need their judgment.
5. Nothing starts from scratch
A document or pull request persists. The agent session that created it should persist too, so someone can return months later and continue with the same context. A new teammate should be able to pick up the work from day one.
Shared agents need shared infrastructure
An agent running on one person's laptop is hard for the rest of the team to access, and closing the lid stops its work. Cloud environments give the team a place to run agents together, at all hours of the day, and to preserve their sessions. They also give the organization a place to control which tools and services agents can reach.
Putting an agent in the cloud is only the start. A shared session needs clear permissions: joining a coworker's chat should not give you access to their private email. Teams also need network restrictions and a record of who asked an agent to act, what it accessed, and what it changed.
Whether you build, buy, or host an open-source solution, keep control of your data and the ability to choose providers. The knowledge, skills, and evals your team builds up are part of its competitive advantage. They should remain yours.
Multiplayer AI for software teams, with Superconductor
Superconductor is the multiplayer AI workspace for teams and coding agents.
This is how we build Superconductor: we talk to coding agents, see live app previews in the cloud, and join each other's sessions to ask questions and make changes. We even do a fair bit of development from our phones. We are building that workflow for the whole team, from the initial request through final review.
Work with the same agent
Teammates can join an implementation chat, see its history, and request changes. Start work from Slack or GitHub and continue with your team in Superconductor. Explore shared agent chat.
Give the whole team a working environment
Agents run in cloud environments with your project's repositories and tools. Shared setup and live app previews let teammates inspect the result without setting up a development environment on their own laptops. See live previews.
Choose agents and measure their work
Run Claude Code, Codex, and other supported agents. Compare implementations and benchmark agents on your own codebase to make choices based on quality, cost, and speed. Learn about benchmarks.
Carry the team's knowledge forward
Keep project instructions and reusable skills alongside your code. Use guided review to understand an implementation, discuss it with the agent, and prepare it for human approval. Explore guided review.
Watch the Multiplayer AI Manifesto
Superconductor cofounder Sergey Karayev walks through the ideas behind multiplayer AI in this presentation on The Full Stack.
Read the Multiplayer AI Manifesto.
Build together with your team and AI agents
Bring a real task to Superconductor. Invite a teammate, work with an agent, and review the result together.