Delta is Zed’s new app built around what it calls DeltaDB, a system for storing coding-agent conversations, related files, plans, and diffs as first-class artifacts rather than disposable chat logs. The pitch is an “agent-native” workspace where you can comment directly on parts of an agent transcript, inspect generated files alongside the conversation, and collaborate with other people in the same session. That puts it closer to Codex-style agent UIs than to a traditional editor, even though Zed says the underlying ideas will eventually flow back into Zed itself.
The reaction landed in two very different places. The strongest positive signal was not “multiplayer coding” at all. It was precise feedback on long agent outputs. People who already work this way said inline comments on plans, diffs, and transcript snippets solve a real annoyance. Today they end up writing bulky follow-up prompts just to point at one paragraph or one hunk. Several commenters said that interaction model already feels valuable in tools like the Codex desktop app, and Delta looks compelling because it extends that pattern while working across model providers and existing agent harnesses. In other words, the practical wedge here is better control of agent output, not shared editing for humans.
Most skepticism focused on everything else. A lot of people do not buy multiplayer development in an editor or collaborative agent chat as a mainstream need. They see version control, pull requests, docs, and screen sharing as good enough for the rare debugging or mentoring case. Others pushed harder on the archival idea itself. Saving every agent conversation as a durable project record struck many as a bad substitute for concise docs or architecture decision records, and some saw it as a privacy and culture problem waiting to happen. The broader mood was that Zed may be chasing the current “multiplayer AI” wave while leaving core editor polish behind. Even commenters who liked the interaction ideas questioned whether Delta’s product surface is defensible, since competitors can copy the UI patterns and some already have pieces of them.
If you use coding agents heavily, watch the emerging UI pattern around inline annotations on agent output and diffs. If you are evaluating tools for a team, treat persistent chat history and multiplayer features as governance and product-risk decisions, not just convenience features.
Mostly skeptical. People liked the inline annotation workflow for steering coding agents, but many doubted the need for multiplayer chat around code, worried that archived transcripts would replace real documentation, and complained that Zed is prioritizing AI products over finishing the editor.
Key insights
01
Inline annotations are the real feature
Direct comments on specific transcript spans, plans, and diffs solve a concrete pain in agent workflows. Instead of writing another long prompt just to point at one bad paragraph or one mistaken hunk, you mark the exact spot and move on. That pattern already exists in plannotator and the Codex desktop app, which makes Delta feel less like a moonshot and more like a strong implementation of an emerging interface standard.
If your team uses coding agents, test tools on how precisely they let users correct output. Better annotation UX may improve throughput more than model upgrades do.
The credible use case for multiplayer editing was not joint authorship of the same file. It was mentoring, debugging, and quick handoffs where control can switch fluidly without wrestling with screen-share ownership. That narrows the value proposition a lot. Delta does not need to prove that people want pair programming all day. It needs to prove that this is better than Zoom, Live Share, or passing a branch around in the moments where collaboration actually happens.
Evaluate multiplayer code tools against support workflows, not against solo coding. If they do not clearly beat screen sharing for debugging and coaching, the feature will stay a demo.
The best case for saving agent conversations is not reading them end to end later. It is using them as a searchable memory of abandoned threads, partially finished work, hidden edge cases, and original scope. One commenter building a transcript-search product argued that retrieval over old sessions is already useful for questions like where work stopped, whether a fix was ever completed, or why a path was rejected. That is a different job from docs. It is closer to organizational memory for work that would otherwise vanish.
If you store agent transcripts, design them for retrieval and audit trails, not passive history. The payoff comes from answering future questions, so search quality and project scoping matter more than raw retention.
A few early testers cared less about Delta’s collaboration story than about using one UI across different model providers, API keys, and agent harnesses. For teams that route work by project, budget, or data controls, that is more operationally important than any chat novelty. It also explains why some people prefer Delta over a single-vendor app like Codex even when the interaction patterns are similar.
If you buy agent tooling for an organization, check whether it respects your model routing, billing, and data-boundary setup. Multi-provider support can be a stronger moat than flashy collaboration features.
Preserving meandering agent conversations can become an excuse not to write tight specs, docs, or architecture decision records. Raw transcripts mix signal with drift, and AI-generated summaries often add another layer of verbosity without giving readers the actual decision and rationale they need. That framing cuts against the entire premise that conversation history should become a core project artifact.
Do not let transcript retention displace explicit documentation. If you adopt tools like this, require teams to distill decisions into short durable records.
Making every agent interaction easy to inspect changes the social contract of development. Some commenters saw it less as collaboration and more as a manager’s dream for monitoring prompts, thought process, and work habits. Teams that already collaborate through pull requests may see little upside in exposing the full path taken to get there, especially when many people still work with agents alone.
Treat transcript sharing as a policy issue before you treat it as a product feature. Decide who can access history, how long it is kept, and whether employees can opt out.
A recurring complaint was not about Delta on its own terms but about what it signals. People who came to Zed for a fast editor think the core product still has bugs and rough edges, and each AI-heavy launch reinforces the sense that basic editing polish is no longer the priority. Even commenters open to experimentation saw a trust problem forming between Zed’s original audience and its current roadmap.
If you are a developer-tools founder, product adjacency does not erase product debt. New AI surfaces can be exciting and still weaken your position if core users think the base tool is slipping.
zed.dev