April 2026 Summaries

Nathan Sobo's announcement of Zed reaching version 1.0 highlights the innovative journey of creating a fundamentally better code editor by building it from the ground up using a unique approach akin to video game development, with a focus on performance and collaboration. Unlike its predecessor Atom, which was limited by its web-based framework, Zed is constructed with a custom UI framework, GPUI, in Rust, allowing for unprecedented control over its software stack. This foundation has enabled Zed to support a wide range of programming languages, integrate Git and SSH, and include AI-native features with multiple agent support through the Agent Client Protocol. The release also introduces Zed for Business, providing centralized tools for companies to deploy the editor across teams. The vision for Zed extends beyond version 1.0, with plans to enhance collaborative coding through DeltaDB, a synchronization engine facilitating seamless human and AI interaction at a granular level. This milestone signifies not the completion but a significant progression in Zed's evolution, with continuous weekly updates advancing software craftsmanship.

Jason Lee, a contributor to Zed since February 2024, has played a significant role in the development of gpui, Zed's custom GPU-powered UI framework, by submitting numerous pull requests, particularly for his company Longbridge's desktop client rewrite. Longbridge, a Singapore-based global stock trading brokerage, opted for gpui to address performance issues with their existing Electron app, leading Jason to contribute essential improvements and create the gpui-component library, which provides a comprehensive set of UI widgets. His efforts have not only enhanced gpui but also expanded its ecosystem, making it a popular choice for projects seeking responsive and native software solutions.

Zed has introduced a new feature allowing the orchestration of multiple agents running in parallel within the same window, facilitated by the Threads Sidebar, which enables users to control and monitor threads by project. This feature operates smoothly at 120 fps and is open-source, offering a new default layout with threads docked on the left, enhancing the organization as users manage various projects and agents simultaneously. The concept of "agentic engineering," which combines human craftsmanship with AI tools, is highlighted as a key principle, aiming to improve software quality while maintaining user involvement. The latest Zed release includes these updates, allowing users to customize their experience and manage workflows more effectively, with the Threads Sidebar providing instant access to thread operations. Additionally, Zed's focus on blending AI and human input is part of a broader trend in software engineering, emphasizing the creation of reliable and adaptable systems.

Zed has introduced Agent Metrics, a public tool providing weekly insights into the adoption and performance of AI agents within its platform, allowing users to compare session counts, turn volumes, and response time distributions across different agents. This tool offers a unique perspective as it includes data from third-party agents supported through the open Agent Client Protocol, reflecting trends specific to Zed's user base from those who have opted into data collection. Key observations include the notable increase in latency for the claude-sonnet-4-6 model over a three-week period, and the dominance of Zed and Claude agents in session and depth metrics, respectively, within the platform's 15.4 million turns across 2 million sessions over the past 90 days. Additionally, analysis of the GPT-5 variants reveals a dynamic and rapidly evolving landscape, highlighting distinct naming patterns and fast-paced updates, which suggest multiple active development tracks. Despite these insights, the tool does not assess the quality of outcomes generated by different agents, as it focuses on usage behavior and response times, rather than the effectiveness of the agents in solving complex problems.

Zeta2, the newly released edit prediction model, offers improved performance by predicting user edits based on preceding activity and allowing acceptance via tab. The model boasts a 30% better acceptance rate and faster responses, achieved through enhanced data intake and knowledge distillation techniques, where a "teacher" model generates training data for a "student" model. Initial training relied on synthetic examples from GitHub commits, but real-world data from user edits provided a more accurate training set. A key challenge was addressing the "reversal problem," where the model erroneously treated user inputs as mistakes, which required refining the teacher's prompt. The base model was switched from Qwen 2.5 Coder to Seed Coder, resulting in better performance metrics. The development process emphasized iterative improvements in data quality, teacher models, and evaluation signals, and included rigorous testing and gradual rollout to ensure reliability. The team plans to continue refining the model and is seeking new team members to expand their efforts.