Key Moments
Zed Agents — with Zed Cofounders Nathan Sobo & Antonio Scandurra
Latent Space PodcastKey Moments
Zed's co-founders discuss their high-performance code editor and its new AI agent features.
Key Insights
Zed is a code editor built from scratch for extreme performance, running at 120 FPS by leveraging graphics hardware.
The editor prioritizes native performance and sophisticated text primitives over web technologies like Electron.
Zed was designed for real-time collaboration from the ground up, utilizing CRDTs for seamless co-editing.
The 'AI era' of Zed integrates LLMs as collaborators, building upon its performance and collaborative foundations.
Zeta introduced 'edit prediction,' generating non-local code suggestions, and Z Agents further expand AI capabilities.
The company focuses on building a superior user experience for AI agents by deeply integrating them into the editor, rather than just relying on LLM tool-calling capabilities.
REDISCOVERING CODE EDITOR FUNDAMENTALS WITH ZED
Zed is being developed as the 'world's greatest code editor,' driven by a commitment to exceptional performance and user experience. Unlike editors built on web technologies like Electron, which can impose limitations on rendering quality and speed, Zed is engineered from the ground up. It achieves a fluid 120 frames per second by running code on graphics hardware, akin to a video game. This approach allows for sophisticated text primitives, enabling features like nonlinear undo and causal history tracking with CRDTs, and real-time collaborative editing.
EVOLVING FROM COLLABORATION TO THE AI ERA
The journey of Zed began with a focus on building a performant and collaborative editor. Early features included real-time pair programming where collaborators could join sessions directly within Zed, working on the same project simultaneously. This collaborative foundation proved instrumental as Zed transitioned into the AI era. The editor's inherent performance and collaborative architecture naturally lend themselves to integrating AI, allowing LLMs to function as another type of collaborator within the development environment. This deep integration is a core part of Zed's strategy.
ZETA AND THE ADVENT OF EDIT PREDICTION
Zed's AI initiatives began with ZI, an integration for working with LLMs directly within the editor. This evolved into Zeta, a fine-tuned model focused on 'edit prediction.' Unlike traditional suggestions that appear at the cursor, Zeta predicts and offers non-local edits across the codebase. This feature allows for insertions and edits at various locations, enhancing the developer's workflow. The development of Zeta involved an iterative process of data collection, supervised fine-tuning, and direct preference optimization, with an open-source approach to the training data and model.
Z AGENTS: INTEGRATING LLMS AS COLLABORATORS
The latest development is Z Agents, which marks a significant step towards full-blown AI agent integration. The team views LLMs as collaborators, similar to human teammates, and aims to facilitate this collaboration within Zed. While acknowledging the increasing ease of LLM integration due to improved tool-calling capabilities, Zed's core competitive advantage, or 'moat,' lies in its technical excellence and the creation of a smooth, performant user experience. This necessitates building the editor from scratch to deeply embed AI functionalities.
NAVIGATING THE COMPLEXITY OF AGENTIC WORKFLOWS
Implementing AI agents involves a careful process that moves beyond simply calling LLMs in a loop. Zed focuses on creating meaningful user experiences through agent interaction, emphasizing evaluation ('eval') as a crucial part of the development cycle. Unlike traditional deterministic testing, LLM interactions introduce non-determinism, requiring new testing paradigms. The strategy involves building a beautiful UI for surfacing agent intelligence, collecting user interaction data, and optimizing the system for delivering valuable outcomes, ensuring the agent's actions are productive and align with user goals.
ADVANCEMENTS IN COLLABORATIVE AI AND MULTI-BUFFER WORKFLOWS
Zed's CRDT-based collaboration is being leveraged to manage changes from AI agents. A key feature is the 'multibuffer,' a panel that aggregates and displays edits made by agents across multiple files. This allows users to review, accept, or reject changes, and even make modifications themselves within this consolidated view. This approach distinguishes between human and AI contributions, maintaining user control and understanding of the codebase. The goal is to create a curated, agentically-derived working set of the codebase that users can actively manage and interact with.
DEEP SYSTEM INTEGRATION AND FUTURE VISION
Future plans for Zed involve providing agents with deeper system access, including debuggers and Git integrations, to enhance their capabilities. The long-term vision is to evolve towards multi-agent workflows and advanced collaboration patterns. This includes exploring parallel agent execution and sophisticated methods for multiplexing changes from multiple agents. Zed aims to redefine collaboration, moving beyond commit-by-commit interactions to a more fluid keystroke-by-keystroke experience, especially when interacting with AI collaborators. The foundation is a vertically integrated product that allows for a cohesive and responsive AI-enhanced development environment.
MCP INTEGRATION AND EXTENSIBILITY
Zed is simplifying the integration of MCP (Meta Communication Protocol) servers, which are essential for agent functionality. Users can install MCP servers through Zed's extension store, streamlining configuration and setup, especially for servers requiring specific API keys or dependencies. This extension model acts as a 'recipe' for installing and configuring servers like the GitHub MCP server, making them easily accessible within Zed. This approach aims to make it straightforward for users to connect and utilize various MCP servers with Zed agents.
PRICING PHILOSOPHY AND OPEN-SOURCE APPROACH
Zed plans to charge for its agentic services, with options for users to either use Zed as a provider or bring their own API keys for LLMs. A subscription model is proposed, offering a set number of requests per month with additional usage-based fees. This pricing strategy reflects the cost of AI compute while maintaining an open-source philosophy. Users have the freedom to use alternative LLM providers like Ollama or their own keys, aligning with Zed's commitment to user choice and flexibility, even as they introduce paid services for the agent features.
EMERGING PROTOCOLS AND A UNIVERSAL AGENT EXPERIENCE
The conversation touches upon the potential for a new protocol for surfacing AI agentic experiences across different tools. Zed's internal CRDT-based collaboration system could form the basis of such a protocol. While Zed prioritizes a vertically integrated experience to showcase its capabilities, the ultimate goal is to enable a universal agent experience that isn't limited to Zed. The aim is to create a beautiful and effective agent integration that can be adopted and expanded upon by other platforms, fostering innovation in agentic development workflows.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Zed AI Agent Quick Start
Practical takeaways from this episode
Do This
Avoid This
Common Questions
Zed is a new code editor built from the ground up with a focus on extreme performance, achieved by engineering it with Rust and running code directly on graphics hardware, akin to a video game. This approach allows for a more fluid user experience and deeper control over the text rendering and editing capabilities, unlike editors built on web technologies like Electron.
Topics
Mentioned in this video
A platform where Zed is described as being famous, indicating its visibility within the developer community.
A framework spun out from Atom, used as a base for VS Code and other applications, which Zed found limiting for user experience.
Operating system whose raw APIs were utilized by Zed for rendering, aiming for complete control over performance.
Graphics API used by Zed for rendering, allowing direct control and aiming to prevent frame drops.
The base model used by Zed for fine-tuning, which was then subjected to supervised fine-tuning and direct preference optimization.
A code editor with an agent mode, mentioned as a point of comparison or existing player in the code agent space.
A new code editor developed by Zed Cofounders Nathan Sobo and Antonio Scandurra, engineered for high performance and built from scratch in Rust.
The programming language used to engineer the Zed IDE from scratch, contributing to its performance.
A code editor with an agent mode, discussed as having a shrinking 'moat' and integration challenges with LLMs.
An earlier AI integration in Zed, described as strapping a text editor to an LLM, serving as a precursor to more advanced features.
A framework asked about by the host, which the Zed co-founder had not looked into extensively, preferring to avoid web tech.
An LLM mentioned as capable of handling tool calling in a loop for agentic loops.
Mentioned in the context of configuring MCP servers, specifically needing access to list GitHub repos for the GitHub MCP server.
A tool that users can choose to interact with instead of Zed's built-in agent provider, reflecting Zed's philosophy of user freedom.
Conflict-free Replicated Data Types used in Zed to enable real-time collaborative editing and nonlinear undo tracking.
A previous code editor created by the Zed co-founders, from which Electron was spun out.
The core AI feature being discussed, representing full-blown agents working inside Zed, with a launch planned for May 7th.
A code suggestion tool mentioned as a point of comparison for Zed's edit prediction feature, Zeta.
Mentioned in the context of a pain point with other editors where linking projects with it requires a dance to get off tokens.
A group or project whose members are fans of Zed and were involved in the first public demo of MCP.
A fine-tuned model by Zed for 'edit prediction', suggesting non-local edits and insertions beyond the immediate cursor.
A code editor built on Electron, mentioned as an example of an application using web technologies that Zed sought to improve upon.
More from Latent Space
View all 89 summaries
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
72 minCursor's Third Era: Cloud Agents — ft. Sam Whitmore, Jonas Nelle, Cursor
77 minWhy Every Agent Needs a Box — Aaron Levie, Box
42 min⚡️ Polsia: Solo Founder Tiny Team from 0 to 1m ARR in 1 month & the future of Self-Running Companies
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free