A developer who had AI writing 25% of his code a year ago is now letting it handle 90%. He ditched his IDE too. Went back to Vi — a 50-year-old editor.
What Is This?
David Crawshaw is the co-founder of Tailscale. He's building exe.dev, an agent platform, and has been publishing real-world reports on AI coding agents for over a year. His first post got 919 HN points, the second 615, and this third installment hit 223 — making it one of the most honest practical reports in the developer community.
The key insight: agent harnesses (tools) barely changed in a year, but models improved dramatically. Crawshaw's own agent Sketch could do things six months ago that popular agents still can't do today. It's the model that matters, not the tooling.
He dismisses public benchmarks entirely — they've all been gamed to death. Instead, focus on qualitative changes. There was no dramatic GPT-2 moment, but steady improvements have compounded to the point where 90% of code can be delegated.
What's Actually Different?
The most shocking change is the death of IDEs. When Copilot launched in 2021, IDEs felt inevitable — autocomplete and inline edits made typing 50% more efficient. But four years later, in the agent era, IDEs are no longer necessary. Agents just need a terminal and access to your codebase.
| Copilot Era (2021-2024) | Agent Era (2025+) | |
|---|---|---|
| Primary tool | VS Code + Copilot | Terminal + Claude Code/Codex |
| Developer role | Writing code + AI assists | Code review + agent directing |
| AI contribution | 50% typing efficiency boost | Writes 90% of code directly |
| Time split | 50% reading / 50% writing | 95% reading / 5% writing |
| Editor | IDE required | Vi/Neovim is enough |
Another major shift: the sheer number of programs. Ideas that would've ended up in an Apple Note titled "TODO" — forgotten forever — are now actually getting built. Crawshaw says he's having more fun programming than ever because the programs he always wished he had time to write actually exist now.
The HN discussion pushed back hard. One developer said their coworkers produce untested slop with AI and it feels like "being gaslit." Many pointed out that coding speed was never the real bottleneck — PR reviews, CI/CD processes, and IAM permissions are. One comment summed it up: AI reduced a 5-day process to a 4.9-day process.
How to Get Started
- Use frontier models first
Crawshaw's strongest advice. Cheap models teach "wrong lessons." Agent capabilities keep shifting, and you need to know the latest frontier model's limits to judge what to delegate. - Ditch built-in sandboxes, use VMs
Claude Code's constant "may I run cat foo.txt?" prompts kill productivity. Spin up a fresh VM per session and let the agent work without constraints. - Practice reading code, not writing it
The time split has shifted to 95:5. Fast and accurate review of agent-generated code is the new core skill. - Remember: "The best software for an agent is whatever is best for a programmer"
Crawshaw's core philosophy. When every customer has an agent, API and developer experience become the user experience. If you're building a product, start with the API.
Crawshaw's Stripe Sigma Example
Stripe launched a SQL query system (Sigma) with a built-in LLM assistant, but the API endpoint was still in private alpha. So Crawshaw gave his agent three sentences of instructions to build a complete ETL pipeline — Stripe API → local SQLite → custom queries. It solved his problem better than Stripe's own product.
Reality Check from HN
Agent experience varies wildly. Power users like Crawshaw see 10x productivity, but most developers haven't reached that level yet. Without completely redesigning your workflow around agents, just switching tools won't deliver the promised gains.
