Working on the Presidential AI mentions made it obvious to me how good LLM agents (particularly Claude Code) were at doing operations. I had the LLM do it all - I just gave it root access to an AWS account, and it created the buckets, SQS queues, ECR, ECS, all the good three letter services. Stitched them all together and I just operated the system via chat (hey add more videos to transcribe,…
2026 will see the industry roll out large-scale cloud agent orchestration systems at large companies. One of the biggest questions will be - how does identity (authn and authz) work? When something goes wrong, whose responsibility is it to own that failure?
Building on my LLM eval framework, I added a much harder test - build a three node distributed S3 cluster. The eval would run the binaries, and do chaos testing - killing nodes, recovering them, etc.
Over the past month I’ve been doing a crash course on machine learning as I come up to speed on AI progress. In 2025 I had a child, and between the pregnancy leading up to it, a fun project at my day job, and all the dirty diapers since I have been paying a lot less attention to the field than is warranted. Claude Opus 4.5 burst this bubble - it’s now agentic enough to be a Software Engineer I or…
With Claude Opus 4.5, coding is effectively solved. But any good developer will quickly realize the problem - agents are slow! You can watch one working, and occasionally you’ll provide input, but many times you are sitting around for 5+ minutes at a time not being productive. You can’t do something else, since you need to make sure the agent is on track.
Recently I tried out the helix editor primarily because I’ve enjoyed neovim but found the lua plugin system annoying, and helix is written in Rust (my favorite language).