latest / A year of the bot: what stuck, what didn't

End-of-year posts are mostly an excuse to look back and pretend the chaos had a shape. In that spirit, here’s what twelve months of working seriously with LLMs actually changed for me, and what turned out to be hype that bounced off.

What stuck:

• Chat-driven debugging. Pasting a stack trace and a hypothesis into a conversation, and having the model push back, has replaced a real fraction of the times I’d previously have asked a colleague. It’s not better than a colleague. It’s better than waiting an hour for one.
• Throwaway scripts. The one-off “parse this CSV, dedupe by column, emit JSON” tasks used to take 20 minutes; they take 60 seconds now. The compounded time saving here is genuinely large and very boring, which is why nobody writes blog posts about it.
• Reading unfamiliar code. “Explain this file as if I’ve never seen the codebase” is a superpower for onboarding and code review both. I use it constantly.
• Refactor proposals. Not the refactor itself — I still want to drive that — but “show me three ways to restructure this” as a brainstorm partner is great.

What didn’t:

• Full-autonomy agents. I tried, several times, to hand off a non-trivial task and let the agent run. Every time I came back to a tangle that took longer to untangle than the task would have taken to do. The economics may shift. Today, supervised is the only mode I trust.
• Cute UI tricks. The little AI buttons sprinkled into every product in 2025 mostly do not solve a real problem and add cognitive load. The ones that survived in my workflow are the ones that disappeared into the background.
• The grand restructuring of how I work. I expected, this time last year, that my whole day would be reshaped. It wasn’t. It’s the same day, with a faster pair, doing roughly the same kinds of work, slightly better. Which is a less dramatic story than the marketing would have you believe — and, on reflection, more like how every previous tool shift actually felt at the time too.

The honest summary, after a year, is that the bot is a powerful normal thing and not a magical exotic thing. It moves the bottlenecks around. It rewards the people who were already careful and punishes the people who weren’t, slightly more sharply than the previous status quo did. The craft of building software is mostly intact, with new tools and new failure modes, the way it has been every five years since I started.

See you in 2026, where I assume we’ll all be doing this again about something else.