I'll be leading the developer advocacy team in EMEA and APAC so expect to see me doing more stuff with workers, agents, durable objects and other aspects of the Cloudflare Developer Platform: developers.cloudflare.com
If you're at React Advanced London tomorrow then drop by our booth and say "hi."

Bottom line: The future of AI-driven discovery isn't just bigger models—it's smarter inference. By combining LMs with rational planning strategies, we can build agents that ask better questions, make better decisions, and collaborate effectively with humans.

One useful trick to improve answering accuracy is to use code generation. Code grounds reasoning in executable logic, not just vibes.

Gwern wrote about this in 2016: gwern.net/tool-ai There is a disproportionate Return On Agency.

I like the "what" versus "how" distinction since it overlaps with @simonwillison.net who has what (a while loop with tools): simonwillison.net/2025/Sep/18/... and how (agentic loops): simonwillison.net/2025/Sep/30/...

What do you think of Karlsson's perspective: www.henrikkarlsson.xyz/p/agency where agency isn't just feedback or a certain kind of loop but a combination of autonomy and efficacy?

PR here: github.com/bradfitz/liv... but I haven't tested it since I don't currently have a working Linux box. I wonder how much of the testing could be pushed onto the Github CI? If so there are a bunch of performance ideas that could be vibe coded relatively easily.

I'd recommend also having ground truthing mechanisms to avoid ending up in LLM psychosis because it magnified your misconceptions.

I just cloned it and told a coding agent to add a bunch of testing infrastructure and tests to figure out if Livecore works. It also had some ideas for improving STW times.
If it actually builds something that works (whilst I'm sleeping) then I'll send you a (hopefully sensible) PR in the morning.

Which coding agent were you using? Claude Code tends to be quite assertive about its contribution but I don't see any of the usual signs.