The IDE is Dead: A Year of Living with AI Agents
A year ago, we thought copilots were the future. We were wrong—and the speed of that wrongness should terrify and excite you in equal measure.
The Core Insight
David Crawshaw’s latest reflection on AI agents delivers a startling verdict: the IDE, that cornerstone of modern development, has become obsolete in just four years. His journey from “Claude can write a quarter of my code” to “Opus writes nine-tenths of my code” happened in twelve months—a pace that makes Moore’s Law look leisurely.
But here’s what’s really fascinating: the agent harnesses haven’t changed much. This revolution isn’t about better tooling or slicker interfaces. It’s purely about model capability. The benchmarks are worthless (they’ve been “gamed to death”), yet the qualitative improvement is undeniable. We’re watching a paradigm shift happen in real-time, measured not by synthetic tests but by the lived experience of developers.
Why This Matters
Crawshaw’s 95-5 reading-to-writing ratio represents a fundamental inversion of the developer’s role. Where we once spent half our time typing code, we now spend nearly all our time reviewing it. This isn’t incremental change—it’s a complete redefinition of what it means to be a programmer.
The implications cascade outward:
- Local models will eventually win, but using anything other than frontier models today actively harms your learning. You’ll internalize the wrong lessons about what’s possible.
- Built-in sandboxes are broken. The endless “may I run cat foo.txt?” prompts from Claude Code are a symptom of a deeper mismatch between safety theater and practical utility.
- Software itself is the wrong shape. Crawshaw’s example of rebuilding Stripe Sigma in “three sentences” isn’t an edge case—it’s a preview of how every product will be evaluated.
Key Takeaways
- The model is everything right now. Pay for Opus/GPT-7.9-whatever. The cost is temporary; the lessons are permanent.
- Use fresh VMs for agent sandboxes. Stop fighting the built-in restrictions and build your own.
- Build for programmers, everyone wins. In the agent era, every user has access to a programmer. The best product strategy is making programmers happy.
- IDEs peaked in VS6.0, then died in 2026. The tool that seemed inevitable in 2021 is now irrelevant.
- Embrace the joy. More programs exist now because the barrier to creation has collapsed. This is unambiguously wonderful.
Looking Ahead
Crawshaw’s philosophy—”the best software for an agent is whatever is best for a programmer”—may be the most important design principle of the next decade. Product managers have long told engineers “you are not the customer.” That era is ending.
The real question isn’t whether agents will reshape software development. It’s whether you’ll be building the tools that work with them, or scrambling to understand why your carefully crafted product got reimplemented in three sentences by someone’s AI assistant.
We have maybe a few years before local models catch up and the frontier model tax disappears. Until then, pay through the nose and learn what’s actually possible. The alternative is worse than wasted time—it’s learned helplessness about the capabilities of tools that are about to become ubiquitous.
Based on analysis of “Eight more months of agents” by David Crawshaw