May 7, 2025 · 35m · latent-space
Zed Agents — with Zed Cofounders Nathan Sobo & Antonio Scandurra
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Zed co-founders Nathan Sobo and Antonio Scandurra discuss the native Rust architecture behind their GPU-accelerated code editor, demonstrating its new autonomous AI agents, Model Context Protocol integration, and vision for real-time multi-agent software development.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.6 Guest teaching 5.0 Guest disagreement 1.6 The hosts pushing back 2.1
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 11:19 Nathan dismisses the notion that code editor agents lack moats
Nathan directly rejects the viral claim that AI editors have no moats, arguing that integrating tool-calling is easy but raw UI performance requires a ground-up engine.
Hardest push from the hosts ▶ 18:14 Host pushes back on linear agent editing in favor of parallel fan-outThe host presses the founders on why their agent runs sequentially instead of spinning up parallel subagents across files to maximize developer reading speed.
Biggest teaching moment ▶ 0:44 Nathan explains why Electron cannot match native GPU performanceNathan breaks down how web application shells impose hard frame-rate limits and explains why building Zed like a video game engine in Rust unlocks real-time 120 FPS rendering.
The host holds their own ▶ 33:48 Host formulates the need for a standardized Coding Agent ProtocolThe host demonstrates deep ecosystem understanding by suggesting Zed open up its internal CRDT collaboration layer into an open industry standard protocol.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Overcoming Electron: The Architectural Roots of Zed | 3 | 6 | 1 | 0 | Nathan and Antonio explain the architectural fundamentals of Zed, detailing why they moved away from Electron, wrote half a million lines of Rust, and implemented CRDTs from scratch. | |
| Desktop Frameworks, Native Performance, and Developer Tools | 4 | 5 | 2 | 2 | The host asks about alternative frameworks like Tauri and admits feeling like an inferior developer for being content with VS Code. Nathan firmly defends building a custom native engine from bare Metal APIs. | |
| From AI Playgrounds to Zeta Edit Predictions | 4 | 6 | 1 | 1 | Antonio and Nathan educate the host on Zeta's edit predictions model, explaining how fine-tuning Qwen 2.5 Coder enables non-local edits beyond standard inline copilot completions. | |
| Agentic Architecture, Tool Calling, and the Shift to Evals | 5 | 5 | 2 | 2 | The host references Torsten Ball's thesis that AI code editors lack moats. Nathan counters that while LLM tool-calling loops are becoming commoditized, true defensibility lies in vertical performance and deterministic evals. | |
| Live Demo: The New Agent Panel and Multi-Buffer Review | 6 | 4 | 2 | 4 | During the live demonstration of the multi-buffer review flow, the host actively questions whether sequential linear tool-calling should be replaced with parallel fan-out subagents, prompting a thoughtful technical debate. | |
| Pragmatic Code Search and Local Terminal Execution | 5 | 4 | 1 | 2 | The host highlights real-world pain points with competing editors running disconnected subshells and executing uncontrolled brew installs, which the Zed founders validate. | |
| Zed's Open Source Monetization and Pricing Philosophy | 4 | 5 | 3 | 4 | The host questions the scalability of maintaining dedicated repositories for individual MCP server extensions. Nathan and Antonio clarify that the repos merely hold lightweight declarative recipes. | |
| The Future of Multi-Agent Development and Final Thoughts | 6 | 5 | 1 | 2 | Nathan articulates a long-term vision of multi-agent CRDT streams replacing git patch workflows. The host synthesizes this into a proposal for an open Coding Agent Protocol. |
Statements from this episode (16)
Sobo: The Zed editor reliably delivers 120 frames per second
“Zed runs at a 120 frames a second, reliably delivering frames.”
Sobo: Zed engineered an editor from scratch to 500,000 lines of Rust
“And we've essentially engineered an entire editor from an empty repo to up to about over a half a million lines of rust.”
Scandurra: Zed's collaborative architecture and native performance structurally benefit AI
“Now it turns out that like the fact that we're collaborative helps a lot also with AI. The fact that we're performant also helps a lot with AI.”
Sobo: Custom architecture gives Zed control to build an AI-native editor
“Like the fact that we even made the investments necessary to Have enough control to deliver excellent performance and collaboration will also give us the control I think it takes to build sort of the first editor that's truly engineered from the ground up for …”
Sobo: Zed targets shipping its code editor to browser tabs in 2025
“Now we're gonna actually ship it to a browser tab. I'm hoping later this year, we'll see.”
Sobo: Great code editors require low-level native graphics APIs, not web tech
“So I can see other points in the trade-off space making sense for different apps, but for an editor, I don't, I think this is what it takes to build the editor I want to use, put it that way.”
Scandurra: Zed fine-tuned Qwen 2.5 Coder using SFT and DPO for Zeta
“So yeah, we started with Quinn, 2.5 coder B and went through the whole process. Collecting examples, fine tune, doing supervised fine tuning, doing direct preference optimization.”
Sobo: Zed trains edit prediction models on opt-in user open-source data
“Like the rest of Zed fully open source, the data that it's trained on is actually open. We're collecting data from any, if you're editing an open source project in Zed, you can opt in to collect the data.”
Sobo: AI editor moats are shrinking rapidly due to better tool-calling LLMs
“So there's a, there was a lot of like integration work to bridge that gap, but now the LLMs, because they've taken on this tool calling and are getting better at tool calling, the story is the integration story has definitely gotten a lot easier. I think it's …”
Sobo: Zed's competitive moat forces competitors to build an editor from scratch
“I mean, to me, the moat Z's moat is like implementing really nice product experiences with technical excellence and their performance, like just delivering a really great user experience that in order to compete with, you would have to Essentially build your o…”
Sobo: Zed's testing infrastructure is fully deterministic and simulates network scheduling
“Like one of the cool things about the way we've done testing and thus far in Zed is everything is fully deterministic. We even simulate like network interaction and we literally simulate the scheduler such that we can interleave deterministically network messa…”
Scandurra: Claude 3.7 uniquely gathers full context first before executing code edits
“Like, we've seen this pattern in Cloud 3.7, which is kind of different from the other models where, like, it really loves to gather up context first, and then it starts doing edits.”
Sobo: Zed's AI agents will likely launch without dedicated semantic code indexing
“We have code indexing like sort of in flight right now, but I think we're probably gonna launch without it just because we're seeing value being delivered already without it.”
Scandurra: Zed supports installing MCP servers via binaries or its extension store
“Like one is you can sort of change your configuration setting point, you know, point to an MCP binary
and you know, Zed just runs that binary. That's one path. The other path that I'm also really excited about is, like, going through our extension store.”
Sobo: Developers prefer real-time keystroke collaboration with AI agents over other humans
“We wanted to collaborate, not commit by commit, but keystroke by keystroke. And that's kind of a weird thing for human beings to want to do. It seems like in this world, like our team works that way. We work a lot together in real time, but a lot of people, th…”
Sobo: The ecosystem needs open agent APIs, not more VS Code forks
“Something that would make me happy is that instead of one more fork of VS code, people just had a great API for surfacing a great agent experience into a tool that was welcoming to that. We don't need to be sharp elbowed about controlling all of that.”