The layer of the AI stack that evolves fastest – and may have the most impact
One of my most popular posts ever was 35 Thoughts About AGI and 1 About GPT-5, a grab bag of musings about the path to AGI (plus a snarky aside about GPT-5).
Here is a fresh collection of musings, this time about AI agents.
- I had trouble making time to write this: I’m so drawn into using agents that it’s hard to make time to write about agents. This isn’t an isolated phenomenon; many people are tweeting about getting sucked into vibe coding at every available minute.
- This is driven by the astonishing productivity of current AI coding agents in particular, especially when used in ways that play to their strengths. I’m using Claude Code to build a ridiculously ambitious set of productivity tools for my own personal use. A single sub-project involves deep integrations with Gmail, Slack, WhatsApp, Twitter, Signal, SMS, Substack, Pocket Casts, Notion, Google Drive, Google Contacts, Google Calendar, and more. As recently as last year, it would have been insane for me to even contemplate such an undertaking, let alone as a side project. With today’s tools, I knocked off most of the integration work over the course of a weekend. AI models are providing the intelligence the app will need to make use of all this data, and coding agents are handling the tedious integration chores.
- After decades as a prolific coder, I stopped cold in early 2023, leaving me quite rusty. That rust hasn’t been the slightest impediment. In fact, I suspect it might be helpful – the habits I would have had to unlearn have already faded, and it’s been easy for me to slip into the habit of letting the AI write all the code. I am still using my high-level design experience to guide the agents; interestingly, those skills don’t feel rusty at all. I wonder what it is about low-level coding technique vs. high-level design skills that makes the latter easier to retain, and what this might say about which human skills will remain relevant.
