Eight More Months With AI Agents: Why 2026 Changed Everything for Developers

The IDE is dead. Long live Vi.

That’s not hyperbole—it’s the lived experience of David Crawshaw, who just shared his updated perspective on programming with AI agents after eight months of intensive use. And his conclusions are reshaping how we think about the future of software development.

The Core Insight

A year ago, Claude Code could write a quarter of Crawshaw’s code. Today, the latest Opus model writes nine-tenths of it. This isn’t incremental progress—it’s a fundamental shift in how humans and machines collaborate on software.

But here’s what’s truly fascinating: the improvement isn’t coming from better agent harnesses or fancier tooling. It’s all about the model. Crawshaw is blunt about this: “Agent harnesses have not improved much… Right now, it is all about the model.”

The numbers tell a stark story. At a big company, developers typically spend 80% of their time reading code and 20% writing it. At startups, it’s closer to 50-50. Now? Crawshaw reports 95-5—and that 5% is mostly reading and adjusting what the AI produces.

Why This Matters

We’re witnessing the death of the IDE as we knew it. In 2021, copilot-style tools made IDEs feel inevitable—the productivity boost was too powerful to ignore. But just four years later, agents have rendered that entire paradigm obsolete.

“The degree of certainty I felt about a copilot future, and the astonishing whiplash as agents gave me a better tool not four years later still surprises me.”

Crawshaw is back on Vi—a text editor turning 50 this year. The only “IDE feature” he still uses is go-to-definition, which neovim handles fine. Everything else? The agent handles it.

This has profound implications for the tools industry. Every SaaS product, every developer tool, every cloud service is now potentially the wrong shape. Crawshaw built his own Stripe analytics solution—implementing their entire Sigma product for his use case—by typing three sentences.

Key Takeaways

• Use frontier models or don’t bother. Cheap models and local alternatives teach you the wrong lessons about what’s possible. Pay for Opus or GPT-7.9 until local models catch up.

• Built-in agent sandboxes are broken. The constant “may I run cat foo.txt?” permission prompts are a nightmare. Solution: use a fresh VM and turn off the sandbox.

• Software needs redesigning. Most software is “the wrong shape” for an agent-first world. Products need to be built for programmers because every customer now has an agent that can write code against your product.

• Joy is back. Despite valid fears about societal impacts, Crawshaw reports “having more fun programming than I ever have” because so many previously-impossible projects now exist.

Looking Ahead

The philosophical shift here is profound: the best software for an agent is whatever is best for a programmer. For decades, product managers have told engineers “you are not the customer.” That’s been completely inverted.

We’re entering an era where customers don’t need polished UIs—they need solid APIs and good documentation because their agents will build custom interfaces for them. The companies that understand this shift will thrive. The ones that keep building traditional SaaS interfaces may find themselves bypassed entirely.

The future isn’t about making tools easier to use. It’s about making them easier to automate.

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