Why IDEs Won't Die in the Age of AI Coding: Zed Founder Nathan Sobo

Core Thesis: The Persistence of Visual IDEs

• Nathan Sobo argues against the "death of the IDE" narrative, contending that source code is a human-readable language requiring visual interfaces for context and comprehension, even as AI agents generate edits.
• He cites Harold Abelson's maxim that "programs should be written for people to read and only incidentally for machines to execute" to justify the continued necessity of navigating raw source code.
• Sobo asserts that while terminal-based AI tools are effective for "vibe coding," they fail at deep code review because they offer only tiny excerpts (e.g., 10 lines) rather than the holistic context needed to verify agent changes.
• He predicts that developers heavily using terminal AI will likely still run a visual editor alongside for inspection, as source code remains the most efficient language for expressing precise, abstract processes.

Zed's Technical Foundation and Performance

• Zed is a modern IDE written in Rust, currently used by approximately 170,000 active developers.
• The architecture was chosen after Sobo reached a performance ceiling with Electron (the web-technology foundation used for GitHub's Atom), which he deemed unsuitable for the zero-perceptible-lag requirements of a professional tool.
• Zed achieves high performance by organizing the entire application around rendering data to shaders on the GPU, similar to how video games render frames.
• The primary design pillars were eliminating web-technology latency and enabling a new model of synchronous, real-time collaboration where team members share the same codebase simultaneously.

Evolution of Collaboration and the ACP

• Zed introduced the Agent Client Protocol (ACP), an open standard designed to connect external AI agents to Zed's interface, mirroring the impact of the Language Server Protocol (LSP).
• The goal is to create an "IDE Switzerland" where multiple competing agents (from Anthropic, Google, JetBrains, etc.) can interact with a single, high-performance UI.
• Sobo notes that JetBrains has integrated with ACP, signaling cross-industry acceptance of the protocol despite competitive tensions.
• The vision extends beyond human-AI collaboration to treat conversations, edits, and code as a unified metadata backbone, allowing context to be permanently anchored to specific lines of code rather than ephemeral snapshots.

LLM Capabilities and Limitations

• Sobo's experience indicates LLMs excel at tasks within the "distribution of their training data," such as generating boilerplate, implementing known patterns (e.g., Tailwind CSS macros in Rust), or acting as a "knowledge extruder" for standard libraries.
• Conversely, LLMs struggle with out-of-distribution problems where the primary challenge is system design and holding complex constraints in memory, rather than just writing syntax (e.g., developing the "delta_db" fine-grained tracking system).
• He emphasizes that the limitation often lies in the "thinking" required to define the problem rather than the coding itself, noting that LLMs are not yet capable of independently solving high-level architectural constraints.
• Sobo views the current state of AI coding as "copy-paste on steroids" for known patterns, while human oversight remains critical for novel problem-solving and avoiding "slop."

Future Interface Vision: The Conversation as an Editor

• The future IDE interface aims to evolve from a chat window into a document-like editing surface where conversations are linear logs of code evolution.
• Proposed features include the ability to edit code snippets directly within a conversation log, with changes synchronizing back to the source files in real time.
• Zed is developing "multi-buffer" capabilities to aggregate scattered files from different parts of a codebase into a single editing context, facilitating complex agent instructions.
• The system aims to replace "snapshot-based" Git workflows (push -> review -> pull) with continuous, fine-grained tracking of every keystroke and edit, creating a permanent history of agent interactions.
• Current adoption metrics show that roughly 50% of Zed's users utilize edit prediction (autocomplete), while approximately 25% are actively using agentic editing features.

User Demographics and Strategy

• Zed currently attracts "hardcore" engineers and seasoned developers who prioritize tactile responsiveness and performance over feature bloat.
• Sobo explicitly avoids the "vibe coding boss" archetype, warning against the risk of generating unmanageable code for junior teams; he advocates for "vibe coding" only when the user has a clear vision and can direct the agent effectively.
• The roadmap includes expanding the tool's appeal to non-expert users (e.g., domain experts like dentists) by improving the user experience, though the immediate focus remains on elite developer needs.
• The team is actively working on "phase two" of their vision: a fine-grained tracking mechanism that acts as a "commit on every keystroke" to permanently anchor feedback and context to the code.