The Death of the IDE: How Eight Months Changed Everything About AI-Assisted Coding

Remember when everyone was certain that AI-powered IDEs would dominate the future of programming? David Crawshaw does—and he just wrote the obituary for that era.

The Core Insight

Crawshaw, who has been documenting his AI programming journey for over a year, drops a bombshell: he no longer uses an IDE. After Copilot seemed to cement the IDE’s inevitability in 2021, agent-based tools have completely reversed that trajectory. Today, he’s back on Vi—a text editor turning 50 this year.

The whiplash is real. In just eight more months of working with agents, his code production ratio has shifted dramatically:

• At a big company: 80-20 reading to writing code
• At a startup (pre-AI): 50-50
• Now with agents: 95-5

That’s not a typo. Ninety-five percent of his programming time is now spent reading code that agents produce.

Why This Matters

This isn’t just one developer’s anecdote. It represents a fundamental shift in how we think about programming tools and productivity:

The Model Is Everything (For Now)

Agent harnesses—the scaffolding around LLMs—haven’t improved much. What has improved dramatically is the underlying models. Crawshaw was blunt: “Using anything other than frontier models is actively harmful.” Not just wasteful—harmful. Because you’re learning the wrong lessons about what AI can actually do.

The Sandbox Problem Is Real

Every major agent tool ships with built-in sandboxes that constantly interrupt your flow with permission requests. Crawshaw’s solution? Just use a fresh VM. Turn off the safety rails entirely—but do it in isolation.

Software Is the Wrong Shape

Here’s the provocative insight: most software products are now poorly designed for an agent-first world. Crawshaw’s example is Stripe Sigma—a fancy SQL query UI with a mediocre built-in LLM. Instead of waiting for their REST API, he had his agent build the entire ETL pipeline from scratch using standard Stripe APIs. Three sentences of instructions. Better results than the actual product.

Key Takeaways

• IDEs peaked in 2021 with Copilot. Agents have now displaced them entirely for some power users
• Frontier models only: Using cheaper models doesn’t save money—it teaches you the wrong limitations
• Built-in sandboxes are broken: Accept the tradeoff and provide your own isolation
• New programming philosophy: “The best software for an agent is whatever is best for a programmer”—because now every customer has an agent that will write code for them
• The fear is valid, but limited: Real concerns exist about labor markets and societal impact, but in the domain of writing code, agents have brought “exploration and joy”

Looking Ahead

Crawshaw is building exe.dev to address the friction he’s experienced—a VM-based environment where agents can run unconstrained. His broader thesis: we need to rebuild software products for a world where customers interact through agents, not UIs.

The provocative conclusion? Build what programmers love, and everyone will follow. The Product Manager’s classic wisdom—”you are not the customer”—has been turned on its head. Every user now has an AI that will write code against your product for them.

Whether you’re building developer tools, APIs, or entire platforms, the question is no longer “how do users interact with this?” It’s “how will their agents interact with this?”

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