Delta (Zed) — multiplayer for the AI age

Zed's new multiplayer development environment where the unit of work is a shared thread: agent conversation and code changes live together, and teammates can join the same thread live.

Links

• Website: https://delta.dev/ (also https://delta.zed.dev)
• Docs: https://delta.zed.dev/docs
• Status: public beta (from 2026-09-16)

Rufus's note (2026-10-05): I've been looking for how to collaborate on an AI thread the same way we could with code via VS Code Live Share etc. Multiplayer for the AI age … and Zed have built it. (I went looking for this on 27 Sep and found nothing packaged.) — Rufus Pollock

Overview

Delta is built by Zed Industries, the team behind the Zed editor, which was multiplayer from the start. Delta takes that idea and moves it up a level: from sharing a cursor in a file to sharing the whole agent session.

• Collaborative threads — several people can join one thread to watch, steer, review and comment while the agent works.
• Conversation anchored to code — every change stays linked to the conversation and decisions that produced it. As one early user puts it: "Git remembers the commit. This remembers everything between them, including the conversation that wrote the code."
• DeltaDB — the version-control layer underneath records work as fine-grained deltas, not snapshots, so comments and references stay attached to specific changes as the code moves.
• In-context review — comments stay anchored to lines and follow the code as it evolves.
• Desktop, web and cloud runners, all looking at the same synced thread.

Why it matters

Today almost all agent work is single-player. One person sits with Claude Code or Codex in their own terminal; the conversation, where most of the reasoning now happens, stays on their laptop and disappears. What reaches the team is a diff and a PR description written after the fact. Code review then has to rebuild the intent that was in the thread all along.

Delta treats the thread as a first-class shared artifact, not a private chat log. That changes three things:

1. Pairing comes back. You can pair with a colleague and an agent at once: one person drives, another watches and steps in when the agent goes off track. This is the Live Share / Google Docs moment for agentic work.
2. Provenance. "Why is this code like this?" can be answered by the conversation that produced it, not by archaeology through commit messages.
3. Version control built for agents. Git was designed for humans making a few deliberate commits. Agents make many small edits quickly. DeltaDB's bet is that the agent era needs a finer-grained history, where intent sits alongside the change.

Our take

The direction is right, and it is the missing piece in the moc-agent-workbenches category: most workbenches help one person run many agents, but Delta helps many people share one agent session. Expect the big labs and IDEs to converge here: shared and hand-off-able sessions are a natural next step for Claude Code, Codex and Cursor.

Open questions:

• Lock-in. Delta runs on its own VCS layer. How cleanly does it round-trip to Git and GitHub, where everyone else still lives?
• Which agents? Does it work with whatever agent you already use (Claude Code, Codex, ACP agents), or only with Zed's own?
• Beyond code. The same need exists for non-code AI work (writing, research, planning), where threads are even more of the output. Delta is code-first; the general "multiplayer AI thread" is still open.
• Noise. Full conversation history is valuable but long. Its usefulness will depend on good summarisation and navigation, not just storage.

Related

• Is there a tool for two people to drive one AI agent session together, live? — my 27 Sep search for exactly this, a week before finding Delta
• moc-agent-workbenches
• claude-projects