Nathan Sobo's Zed takes aim at pull requests with DeltaDB

Why it matters

Zed is betting that AI coding shifts the durable record of software from commits to conversations. If DeltaDB works, the editor becomes more than a place to write code; it becomes the system that preserves who or what changed the code, when, and why.

Nathan Sobo (@nathansobo)'s Zed is introducing DeltaDB, a version-control system meant to capture the edits and agent conversations that happen before code becomes a Git commit, according to a Zed blog post.

The move fits the origin story Zed has been telling since its 2021 founding: the editor was built not just to be faster than Electron-based tools, but to make live collaboration part of the coding surface itself. In Zed's 1.0 post, Sobo wrote that Zed's team previously built Atom, a Chromium-based editor that spawned Electron, before starting over with a Rust UI framework called GPUI. DeltaDB is the next piece of that rewrite: not another AI sidebar, but an attempt to change where software collaboration is recorded.

Zed's argument is direct. Pull requests and review comments attach discussion to code after a developer has committed and pushed it. That model worked tolerably when humans wrote code locally and summarized their decisions after the fact. It breaks down when AI agents are producing larger chunks of code through long-running conversations, because the conversation itself becomes part of the engineering record.

"Increasingly, the conversation that generates the code is becoming the true source of our software," Zed wrote in the post. "Git, organized around discrete commits, was never designed to support this."

What DeltaDB is trying to capture

DeltaDB breaks work into fine-grained "deltas" rather than waiting for commit-level snapshots. Zed says each delta gets a stable identity, allowing a developer to point to a particular moment in a file's evolution even as the file keeps changing.

That matters because agentic coding creates two linked artifacts: the message that asked for a change and the edit that resulted from it. Zed says DeltaDB records those side by side, so a developer can jump from a line in an old agent conversation to the code as it looked then, or to the code as it stands later.

Zed also says DeltaDB embeds conflict-free replicated worktrees, allowing multiple people and agents to edit the same files across machines at the same time. The post stresses that these remain real files: agents can work in them through a terminal, and developers can mount the worktree to disk for other tools.

That is a more ambitious claim than adding AI to an editor. Zed is trying to own the collaboration layer before Git sees the work.

The GitHub critique is about timing

Zed is not saying Git disappears. The post explicitly leaves Git and CI in place for checks and for connecting code to the broader ecosystem. The challenge is aimed at the point in the workflow where collaboration currently becomes visible.

On GitHub, review usually begins after a commit, branch, push, and pull request. Zed's critique is that the highest-value discussion often happened earlier, while the worktree was changing. In the human-only version of this workflow, that discussion might have been a screen share, chat, or pair-programming session. In the agent version, it is a prompt trail, a sequence of edits, and the reasoning context that led to them.

The strategic bet is that the editor becomes the system of record for that pre-commit work. Zed has already been pushing in that direction with real-time collaboration, remote development, multibuffer editing, native Git support, and agentic editing features including Parallel Agents. DeltaDB extends that logic into version control itself.

The timing is not accidental. AI coding tools have made code generation faster, but they have also made provenance messier. A pull request can show what changed. It rarely shows why an agent chose a pattern, what alternatives were discussed, or which earlier agent thread introduced the assumption a later thread is now modifying. Zed wants those links to be addressable objects, not artifacts buried in chat logs.

The beta still leaves commercial questions open

Zed says a DeltaDB beta will be ready "in a few weeks" and is taking signups through the DeltaDB waitlist. The announcement does not specify pricing, hosting, licensing, or how DeltaDB will be packaged with Zed for individual developers versus teams.

Those details matter because version control is infrastructure, not a cosmetic feature. Developers will want to know whether DeltaDB data lives locally, in Zed-hosted infrastructure, or in a customer-controlled environment. Teams will also need to understand how DeltaDB interacts with compliance, retention, code review policy, and existing Git providers.

Zed has signaled that Git remains the bridge to the outside world. That is the pragmatic path: replacing Git is not the near-term wedge. Capturing the work before Git sees it is.

For Sobo and Zed, DeltaDB also sharpens the distinction from the broader AI-editor market. Many products are competing on model access, autocomplete quality, or agent orchestration. Zed is making a deeper infrastructure claim: if software is increasingly made between commits, the winning editor is the one that can preserve that in-between state as a first-class artifact.