How Zed Built An Incredible Agentic Code Editor: A Conversation with Nathan Sobo

How Zed Built An Incredible Agentic Code Editor: A Conversation with Nathan Sobo

Learn about Zed's incredibly high-performance code editor and how they built streaming edits as part of their new agentic editing feature.

Learn about Zed's incredibly high-performance code editor and how they built streaming edits as part of their new agentic editing feature.

In our latest episode of Deployed, we talk with Nathan Sobo, co-founder and CEO of Zed — the high-performance code editor that's been turning heads in the developer community. It’s not just a talk about code editors though! Nathan loves his product, and it’s a great conversation with relevant product design and AI development insights for any team no matter what you're building.

For more than a decade, Nathan’s been focused on building a better IDE. He was the creator of the Atom editor at GitHub, and he started working on Zed years ago (before building new code editors was cool). Now he's building what might be the fastest full-featured code editor available today. It’s written in Rust and even uses the GPU for acceleration and to ensure it’s always a smooth UX.

What makes this conversation particularly valuable for our audience is how the Zed team has integrated AI agents into their editor, creating what feels like the most natural agentic coding experience available. Nathan shares a bunch of insights that will be interesting to anyone who cares about developer experience for sure, and that also apply well beyond code editors for anyone building production-grade AI products.

Check out the full episodes on Spotify, Apple Podcasts, or YouTube. Some of our favorite highlights below.

Automate Your Feedback Loops (Especially for AI)

Nathan's number one piece of advice for AI product builders comes from years of rigorous software development, and it’s the heart of why we’re building Freeplay:

"I can't stress enough the importance of getting rigorous, systematic automated feedback loops in place... it's never been more true than working with this sort of non-deterministic technology. Like, I don't know how you get anywhere without having that kind of infrastructure in place."

In the full episode, Nathan shares how this played out when they implemented streaming edits. Their AI model was generating malformed XML tags about half the time, causing crashes. Through systematic evaluation, they improved from 60% success to 95%+ (which means they still had to make their parser tolerant of edge cases).

"Up until now, almost all of our testing has been very deterministic... But with LLMs, all of that kind of goes out the window. It's like this black box component and there's literally no way to know what it's going to do without actually seeing what it does."

Addressing this challenge with generative AI is fundamental, and it’s part of why so many leaders who come on Deployed say they wish they’d started with their evals, or that evals are the central piece to making their production systems work. 

Incorporating AI Agents Into Your Product UX

One of the most interesting aspects of Zed's AI integration is how naturally it fits into their existing product. This was a bit of good luck, as the result of architectural and design decisions made years before AI agents existed:

"A big thing that we baked in from the beginning was collaborative data structures... That was originally designed for humans, but now we have these handy new human-like entities (that fit right in)... You can follow the agent around... you could stop following and start typing and its edits would cleanly interleave with stuff that you're doing."

This felt like an interesting UX insight for any product teams, and extends well beyond code editors. What would it look like for an agent to show up in your existing product, without needing to create a new UI element? It’s an interesting thought for any product team.

Making AI Compelling for Skeptics

Zed's core community includes a lot of deeply experienced developers, many of whom have been AI-skeptical. They’re writing complicated and high-performance code, and it needs to work.

"(When) we were at RustConf... some people are like, 'There's AI in your editor? I'm not interested'... these developers are working on embedded systems... people writing databases... it's a very kind of hardcore set of engineers."

One of the unlocks Nathan’s seen that has helped build trust for this group: Their team created a "subtle mode" for code completions:

"We have this thing called ‘subtle mode’... we have a prediction for you, but we're going to hold off showing it to you and like breaking your train of thought until you decide that you want it… I saw a lot of kind of AI skeptical people... when we landed the ability to do these co-pilot style edit predictions... even just getting a little bit more efficiency in like a traditional editing flow did make a big difference."

There’s another AI product design principle in here that applies broadly: Giving people access to AI without making it awkward, and giving them the choice for when to use it. As more teams seek to build confidence in AI features among existing users, it seems like a smart pattern consider.


For teams building AI products, Nathan's experience offers valuable lessons: invest in the feedback loop you need to improve product quality from day one, make smart decisions about UX, and more than anything, simply care about product quality. Nathan’s passion for building a perfect product is compelling for any builder.

Want to try Zed yourself? Check it out at zed.dev. And subscribe to Deployed on Spotify, Apple Podcasts, or YouTube for more conversations like this.

Categories

Podcast

Authors

Ian Cairns

Subscribe to our newsletter