Zed's Delta reframes Xanadu's vision for AI coding agents

Computing pioneer Ted Nelson coined the word hypertext in 1965 as part of Project Xanadu, his lifelong effort to build what he called the docuverse: a system that keeps every version of every document, where hypertext links know both their source and destination and quotations are kept by reference rather than by copy, so that any included text keeps its identity and source down to the level of a single span. Nelson called this being xanalogical, defined by two rules: never copy, always reference (a practice he termed transclusion), and never overwrite, always version.

That vision needed unlimited storage and a naming scheme that would never break, and neither existed for the decades Nelson worked on the project. When the web took off in the 1990s, developers built it on simple string links that break when their target moves, pushing the burden of upkeep onto users instead of the system, because that let people ship fast. Xanadu, the post says, faded into computing's most famous vaporware. Wired's account framed Nelson's story as one of mismanagement; the author disagrees, arguing instead that Nelson had simply specified Xanadu decades before the technology to build it existed.

Zed's engineering blog argues that AI coding agents are the user Nelson's design was actually waiting for. Unlike people, the post says, agents can track more layers of context at once, such as the sources behind a quotation, the discussion around it, and the stack trace tied to the exact code that produced an error, and a fragment-based representation keeps each of those references attached losslessly to the same span of text as it changes, letting an agent cite its sources so a human can audit its work. Zed, maker of the open-source Zed code editor, says this thinking sits behind two things it is introducing: Delta, a shared, permanent history of both a coding conversation and the code itself, and DeltaDB, the database underneath it.

On screen, the post says, a file in DeltaDB still looks like an ordinary string of characters, but underneath it is represented as fragments with stable identities. Those identities let the system create anchors: references to a span of text that keep resolving even after the surrounding code has changed, where an ordinary line number only describes where text sits in a single snapshot. Files and symbols that earlier agents repeatedly read, edited or cited can become landmarks for the next agent, resolved against the current code and linked back to the conversation that produced that understanding; by keeping the causal history behind each fragment, meaning which operation produced it and what state it built on, DeltaDB is meant to give a model access to a codebase's provenance and an agent's prior reasoning, not just its current text.

The post credits sixty years of unrelated engineering work for making Delta buildable now, though none of it existed in Nelson's time: Lamport timestamps (1978), which name every operation by every human or agent with an actor and a logical clock; Merkle trees (1979), which Git made ordinary in 2005 and which DeltaDB uses to name every state between commits; CRDTs, formalized in 2011 and, the post says, central to Zed's own work for the past decade, which let several people and agents edit the same worktree from different continents at once without coordinating; storage, whose cost the post puts at tens of thousands of dollars per gigabyte in 1981 versus a penny now; always-on broadband, which overtook dial-up in the mid-2000s; Firecracker-class microVMs (2018), which let Delta spin up an isolated cloud machine for an agent in milliseconds; and Max Brunsfeld's Tree-sitter (2018), fast enough to reparse code as it is typed, which the post pairs with Zed's own GPUI for deriving a live interface from application state as it changes.

The post says Xanadu's other failure was self-inflicted: it refused to interoperate with lesser formats. Delta is built to avoid that by working with the git repository a team already has: every thread, a captured coding session, is also a real git branch, so teammates who never open Delta still see a normal repository, and because the underlying files are real, any tool an agent can use still works with them. Teams not ready to switch can keep mirroring their repo to GitHub, which the post says is Zed's own current approach for its open-source editor during what it calls this transitional period.

The post closes by naming Engelbart alongside Nelson as the two prophets of early computing whose visions, it says, were never shipped whole: from Engelbart computing kept only the mouse (and eventually multiplayer editing), and from Nelson only the word hypertext, in both cases at a shallower depth than their creators intended. The author's hope is for Delta's threads to meet both visions at once, a live session that also leaves behind a permanent, connected record, because in this design the session and the record are meant to be the same object. The post states that every property needed to build Delta and DeltaDB was specified by Nelson before the author was born, and argues Nelson's one blind spot was not imagining an artificial reader as Xanadu's user, unlike science fiction, which already had Asimov's Multivac and Stephenson's Librarian; in the post's own framing, that new kind of user, not any missing technology, was the final piece Xanadu had been waiting for.

Key facts

• Zed's engineering blog introduces Delta, an agent-native version-control and collaboration system built on a new database, DeltaDB, that stores files as fragments with stable identities rather than flat text.
• The post frames AI coding agents as the missing user for Ted Nelson's 1965 Project Xanadu, whose two founding rules were never copy, always reference (transclusion) and never overwrite, always version.
• It credits a stack of technologies matured over sixty years for making Delta buildable now: Lamport timestamps (1978), Merkle trees made ordinary by Git in 2005, CRDTs formalized in 2011, Firecracker-class microVMs (2018), and Max Brunsfeld's Tree-sitter (2018).
• DeltaDB's fragment identities create 'anchors,' references to a span of code that keep resolving after the surrounding code changes, unlike a line number, which describes only one snapshot.
• Every Delta thread is also a real git branch, so teammates who never open Delta still see a normal repository, and Zed says it still mirrors its own open-source editor's repo to GitHub for those not yet using Delta.

Why it matters

Version-control tools were built around how people read code: someone opens a diff, skims a changed function, and relies on memory to fill in what a line number cannot show. The post argues agents work the opposite way. They hold nothing in memory between turns and, in practice, read far more of a codebase's history than any person would, so details that are fine for a human to lose, the reasoning behind a change, the discussion that shaped it, which version of a quoted fragment is meant, become a real gap for an agent. Ted Nelson specified exactly this kind of permanent, fully attributed record for his 1965 Project Xanadu and never got it built. The post's claim is that missing technology blocked Xanadu for decades, but its final missing piece was a user who could let the system handle the bookkeeping, and that AI agents are finally that user.

Who it affects

Directly: developers who use or are considering Zed's code editor, and specifically anyone running AI coding agents against a real codebase, since Delta and DeltaDB are pitched as the substrate those agents work on. Indirectly: teams that have not adopted Delta, because the post says every Delta thread is still a real git branch, so a collaborator who never opens Delta just sees an ordinary repository; Zed says it will keep mirroring its own open-source editor's repo to GitHub for contributors not using Delta yet. More broadly, the piece speaks to anyone interested in the history behind Nelson's Project Xanadu and Engelbart's 1968 demo, which it treats as unfinished business now being revisited.

How to use it

Delta is pitched as something that fits into an existing workflow rather than replacing it: a thread, one captured coding session, is simultaneously a real git branch, so a teammate who never opens Delta still sees an ordinary repository, and because the underlying files are real, any tool an agent already relies on keeps working against them. A team not ready to switch can still be reached by mirroring its repo to GitHub, which the post says is exactly Zed's own current approach for its open-source editor during what it calls this transitional period. The post gives no price, version number, or release or availability status for either Delta or DeltaDB, so there is no stated way yet to actually sign up or start using either one.

How solid is it

The claim comes from Zed itself, as a first-person engineering-blog post announcing its own product, not from independent reporting. Its case rests on a historical argument: a list of underlying technologies, Lamport timestamps, Merkle trees and Git, CRDTs, cheaper storage, faster networks, microVMs, Tree-sitter, each dated by the post itself, plus a description of DeltaDB's fragment-and-anchor design. The post gives no benchmark or performance numbers for Delta or DeltaDB itself, and no customer names, case studies, or usage figures, so there is nothing here to check the architecture's real-world behavior against.

Risks and caveats

Every claim about why Xanadu failed and why Delta succeeds where Xanadu did not comes from Zed's own retelling, with the company acting as both narrator and product owner. The framing of Xanadu's downfall as self-inflicted, 'it refused to interoperate with lesser formats,' is the post's own interpretation, offered alongside, not instead of, Wired's mismanagement account, and the post names neither the Wired piece nor a source for how little funding Nelson had. The extended list of enabling technologies reads as persuasive narrative but does not amount to a demonstration that DeltaDB performs at scale, works across a large team, or holds up against real-world edge cases; none of that is tested or shown here.

“Though Xanadu was blocked for decades by missing components, what it needed most was the perfect user.”

— Zed's blog post introducing Delta and DeltaDB