Where Your Coding Agent Actually Lives Now

Simon Willison thinks the busiest surfaces for coding agents have left the IDE. The evidence is mounting — and it changes which tools, and which habits, are worth betting on this year.

A throwaway question from Simon Willison this week named something many developers feel but rarely say out loud: the most-used surfaces for coding agents may no longer be the IDE at all. If he's right, the tools you reach for — and the habits you build around them — are being rewritten underneath you while you're still tabbing through autocomplete.

Willison asked whether "the most popular interface surfaces for coding agents these days no longer live in developer IDEs." He framed it as a question, not a verdict, but the supporting evidence keeps stacking up. Anthropic now ships Claude Code Desktop and Claude Cowork as standalone agent surfaces, and the engineer who leads both walked through his daily workflow this week — notably, much of it happening outside a traditional editor. The center of gravity is drifting from the editor pane toward the terminal and the chat window.

The IDE isn't disappearing, though. A 2026 roundup from The Software Scout still calls VS Code the safest default for most developers as of May 2026, with Cursor's free tier the obvious on-ramp for anyone curious about AI-assisted editing. The honest read: agents that run from a standalone app and AI woven into the editor you already know are operating side by side, not replacing each other yet.

For people actually shipping code, the sharper question is how you drive these agents. Jason Liu is sketching a "micro vs macro" framing — tight, scoped edits versus large autonomous tasks — which maps cleanly onto a working day: micro when you know exactly what you want changed, macro when you're delegating a whole feature. The discipline that makes the macro mode safe is testing. Willison argues he's "firmly on team red/green TDD for agent code," because a test suite is what stops an agent from quietly breaking old features while building new ones — and maintaining those tests is cheap now that the agent writes them.

Underneath the interface debate, the model layer keeps commoditizing. Google's CodeGemma is positioned as a lightweight open model for code completion and generation — a reminder that raw coding capability is increasingly something you can self-host, not only rent from a frontier vendor.

The bottom line: Choose your agent surface deliberately — editor or desktop — but invest in the test suite first, because that's the one thing that makes either one safe to let off the leash.

Sources