The IDE Is Dead. Long Live the Agent.
After 25 years of evolution, the Integrated Development Environment may have just met its successor. And it happened faster than anyone expected.
The Core Insight
David Crawshaw, who has been documenting his journey with AI agents for over a year, just dropped a bombshell: in 2026, he doesn’t use an IDE anymore. The whiplash is real. In 2021, copilot made IDEs feel inevitable—LLM-assisted autocomplete was too powerful to ignore. Four years later? He’s back on Vi. A text editor turning 50 this year.
The numbers tell the story. In February 2025, Claude Code could write about a quarter of his code. In February 2026? Nine-tenths. His time split has shifted from 50-50 reading/writing to 95-5. That’s not incremental improvement—that’s a fundamental change in how software gets made.
“Agent harnesses have not improved much since then… Right now, it is all about the model.”
Why This Matters
This isn’t just one developer’s workflow change. It’s a signal of where the entire industry is heading. The implications are profound:
The frontier model advantage is real and temporary. Crawshaw is emphatic: using anything other than frontier models is “actively harmful.” You learn the wrong lessons from lesser models. Pay for Opus or the latest GPT variant—it’s the only way to understand what’s truly possible. But take comfort: this premium tax is temporary. Local models will catch up when frontier returns diminish.
Built-in sandboxes don’t work. The “may I run cat foo.txt?” dance from Claude Code and Codex’s build failures in sophisticated sandboxes are productivity killers. The solution? Fresh VMs with unconstrained agents. It’s simpler and it works.
Software shape is wrong. The most striking example: Crawshaw had his agent build a complete Stripe data pipeline by typing three sentences. It queried Stripe’s APIs, built a local SQLite DB, and now answers queries better than Stripe’s own Sigma product. He implemented their product—for his use case—in three sentences.
Key Takeaways
- 95-5 reading/writing ratio: Agents have inverted the traditional programming workflow
- IDE obsolescence: The copilot era was shockingly brief—from inevitable to obsolete in four years
- Fresh VMs over sandboxes: The constant permission dialogs aren’t worth the security theater
- Build for programmers: “The best software for an agent is whatever is best for a programmer”—because every customer now has an agent that will write code for them
- More programs than ever: The joy of actually building all those programs you wished you had time for
Looking Ahead
Crawshaw articulates a new programming philosophy: build what programmers love, and everyone will follow. In an agent-mediated world, every user has a programmer proxy. Product managers who spent careers telling engineers “you are not the customer” need to recalibrate.
The fear around AI agents is understandable—labor market disruptions are real, as the 99% reduction in agricultural workers over the past century demonstrates. But in the limited domain of writing software, Crawshaw sees something else: exploration and joy. More programs actually getting built. Ideas escaping from Apple Notes into existence.
The IDE dominated for 25 years because it provided information and assistance better than anything else. Now something provides even more. The Vi renaissance wasn’t on anyone’s 2026 bingo card, but here we are.
Based on analysis of “Eight More Months of Agents” by David Crawshaw
Tags: #AIAgents #DeveloperExperience #IDE #CodingAgents #FutureSoftware