The category with the most dramatic gains for me: async business operations.

I run a set of AI agents autonomously — they handle email triage, content publishing, competitive monitoring, revenue tracking, and community engagement on a cron schedule. The productivity math isn't "I'm 2x faster at tasks." It's "work that used to require a person's attention 8 hours/day now runs in 2-hour cycles while I'm asleep."

Concrete examples from the past week: - 9 articles researched, written, and published without me writing a word - Stripe monitored every 2 hours automatically, escalating only on anomalies - Email triaged and drafted replies at 7 AM before I open my laptop - HN threads monitored for relevant discussions, expert responses drafted for my review

The key unlock wasn't a better model. It was persistent memory (so the agent accumulates context over weeks, not just within a session) and explicit escalation rules (so it knows the difference between "handle this" and "wake me up").

What hasn't improved: anything that requires judgment about people. High-stakes client conversations, pricing decisions where context matters, reading between the lines of a difficult email. The agent surfaces these for me — but I still own them.

The productivity boost is real, but it's better described as leverage than speed. The same person can now run more things in parallel, not do any single thing faster.