Building an ACP-Compatible Agent Live — Bennet Fenner, Zed
AI Engineer Europe 2026 · 18:19
Developer tools and AI coding
Zed builds an open-source code editor for developers working with teammates and AI agents on macOS, Windows and Linux. Zed combines collaborative editing, debugging, Git operations and edit prediction, with support for running agents in parallel. Developers can connect agents through Agent Client Protocol and delegate changes while reviewing their progress. Its separate application, Delta, entered private beta in 2026, letting teams work with agents and review their output while keeping conversations connected to evolving code.
Founded in 2021, Zed is led by CEO and co-founder Nathan Sobo, alongside co-founders Antonio Scandurra and Max Brunsfeld. All three previously worked on Atom; Brunsfeld also created Tree-sitter, an incremental parsing framework. Zed’s engineering approach centers on GPUI, its Rust UI framework organized around GPU rendering. Delta uses DeltaDB to synchronize conversations and working code in real time, capturing edits between Git commits without requiring teammates to abandon their existing Git repository.
In 2026, the company reported hundreds of thousands of daily users of its editor. Zed offers an optional paid service alongside the open-source product. It announced a $32 million Series B led by Sequoia Capital in 2025, bringing total funding to more than $42 million.
AI Engineer Europe 2026 · 18:19
AI Engineer Europe 2026 · 10:50
AI Engineer World's Fair 2025 · 14:50
Affiliations reflect their AIE appearances, not necessarily current employment.
Start here to follow a minimal TypeScript agent backed by the Anthropic API, including its agent loop, filesystem tools, and integration with Zed.
Bennet FennerAI Engineer Europe 2026
Learn how replayable concurrency simulations and SWE-bench-like evaluations test different aspects of an editor with agentic coding.
Nathan SoboAI Engineer World's Fair 2025
Learn how cursor context, recent edits, definitions, and diagnostics inform predictions, and how settled-state capture and Levenshtein distance support evaluation.
Ben KunkleAI Engineer Europe 2026
Bennet Fenner’s live implementation makes ACP concrete through JSON-RPC communication, capability negotiation, session state, cancellation, and streamed output.
Nathan Sobo combines end-to-end agent evaluations with focused regression tests, addressing failures such as streaming-diff errors, indentation mismatches, and model-specific escaping.
Ben Kunkle describes repairing teacher outputs and using an increasingly capable student to reduce expensive teacher sampling. Production experiments track acceptance rates and latency.
Affiliations reflect each recorded session, not necessarily current employment.