The Death of the IDE: How AI Agents Killed a 25-Year Assumption

In 2021, the IDE had won. Copilot made auto-complete so powerful that fighting with environment integration seemed worth it. Four years later, David Crawshaw—founder at Tailscale and Sketch—is back on Vi.

Vi is turning 50 this year.

The Core Insight

Crawshaw’s latest update on AI agent development delivers a stunning reversal: agent harnesses haven’t improved much in a year, but models have improved so dramatically that everything else became irrelevant.

The numbers are striking:
– February 2025: Claude Code could write about 25% of his code
– February 2026: Latest Opus writes 90% of his code
– His read-to-write ratio shifted from 50-50 to 95-5

This isn’t incremental progress—it’s a phase transition. And the implications ripple through everything from tooling choices to product philosophy.

“The only IDE-like feature I use today is go-to-def, which neovim is capable of with little configuration. So here I am, 2026, and I am back on Vi.”

The IDE, that seemingly inevitable destination of modern development, got routed around in less than five years.

Why This Matters

Crawshaw surfaces several uncomfortable truths for the broader developer ecosystem:

Frontier models or nothing. Using cheaper models like Sonnet or local alternatives isn’t just suboptimal—it’s “actively harmful” because you learn the wrong lessons about what AI can do. The frontier keeps moving; your intuitions need to keep up.

Built-in sandboxes are theater. The constant “may I run cat foo.txt?” permission prompts from Claude Code and Codex’s sandbox failures make official safety measures impractical. Crawshaw’s solution: fresh VMs with unconstrained agents.

Software is the wrong shape. His Stripe example is devastating: he needed SQL access to his Stripe data, but Stripe launched a fancy UI with an “integrated helper” before their API. So his agent did ETL from scratch—querying everything via standard APIs, building a local SQLite DB, and running queries better than Stripe’s own product.

“I implemented that entire Stripe product (as it relates to me) by typing three sentences.”

Key Takeaways

• Ignore public benchmarks—they’ve been “gamed to death.” Trust qualitative experience with frontier models.
• Agent harness innovation is stalled—model improvements dominate everything right now.
• More programs exist because one-liner ideas actually get implemented. This is bringing “joy” back to programming.
• Anti-LLM arguments have lost him—they sound like “someone saying power tools should be outlawed in carpentry.”
• Build for programmers, build for everyone—every customer now has an agent that will write code against your product.

Looking Ahead

Crawshaw’s new programming philosophy is worth sitting with: “The best software for an agent is whatever is best for a programmer.”

Product managers have long told engineers “you are not the customer.” That’s been inverted. Every customer now carries a coding agent, so building what programmers love becomes building what everyone can use.

This has profound implications for API design, documentation quality, and the entire product surface that software exposes to the world. The era of fancy UIs masking poor APIs is ending—agents expose those gaps immediately.

The most honest moment in the piece: Crawshaw admits he’s “extremely out of touch with anti-LLM arguments.” After pushing agents to their limits daily and still finding transformative value, the fear-based takes just don’t compute anymore.

For developers wrestling with whether to invest deeply in AI tooling, this is the signal: the people who’ve gone furthest aren’t looking back. They’re too busy writing three-sentence prompts that replace entire products.

Based on analysis of “Eight more months of agents” by David Crawshaw