Sunil Pai is a principal systems engineer at Cloudflare and the founder of PartyKit, the open-source real-time collaboration platform Cloudflare acquired in 2024. He builds infrastructure that gives applications—and increasingly AI agents—persistent state, shared connections, and tightly controlled access to executable code.
Pai worked on Facebook’s React team, helped develop the JavaScript stack at Indian e-commerce company Myntra, and created glamor, a CSS-in-JavaScript library. At Cloudflare, he worked on developer tooling including Wrangler; work on multiplayer applications at Meta’s Oculus division deepened his interest in collaborative software.
He founded PartyKit to simplify the distributed-systems challenges behind shared documents, whiteboards, games, and synchronized interfaces. Developed initially alongside tldraw, the platform used Cloudflare Durable Objects to manage persistent rooms, WebSocket connections, and shared state. PartyKit raised $2.5 million in pre-seed funding in 2023, when Pai was its chief executive.
Cloudflare acquired PartyKit in April 2024, bringing Pai back to the company. The same real-time multiplayer infrastructure now informs his work on AI systems that maintain identities, reconnect across devices, and coordinate activity over time.
What he is building
- Code Mode: Pai co-developed an executable-code interface for AI agents that replaces repeated individual tool calls with generated JavaScript against typed APIs. A single program can discover endpoints, sequence requests, loop through results, and preserve intermediate state. He attributes Cloudflare’s compact search-and-execute design to colleague Matt Carey.
- Capability-based security: With Dynamic Workers, generated programs execute in lightweight isolates without ambient privileges or unrestricted network access. Developers explicitly grant narrowly scoped capabilities and can inspect what the generated code actually does.
- Stateful collaborative agents: Through Cloudflare’s Agents SDK, Pai applies Durable Objects to persistent agent identities, resumable streams, scheduling, and synchronization across browsers and devices. His aim is to make AI interactions collaborative without forcing application developers to reconstruct distributed coordination themselves.
- Project Think: Co-developed with Kate Reznykova, Project Think extends agent infrastructure toward durable execution, persistent workspaces, sandboxed code, delegated sub-agents, and reusable extensions.
Pai also sees generated code enabling applications to adapt to individual users: agents can inspect existing software state, assemble specialized workflows, and produce interfaces tailored to particular tasks. His skepticism remains sharply practical: in a widely shared public satire, he mocked autonomous software factories obsessed with orchestration and feedback loops but unable to articulate what they actually produce.
