I'm finding that the bottleneck with agents quickly shifts to testing, especially if you run many agents at once. The better & more automated your testing, the faster you can go

I’m seeing so many more execs and leaders getting into the weeds and building with their own two hands compared to ~4 years ago when managers tended to stay away from the details. Such is the pull of a generational technology shift

Coding agents feel goofy sometimes bc they carry almost zero context between sessions. There's compaction methods, but atm they're not there yet. So you have to explicitly document what feels like any relevant context, even things like how many engineers work in the codebase

when openai, anthropic, and google all drop in the same week

let's just hope Google doesn't build a misaligned AI or I am toast

Tried playwright MCP with Claude Code to get it to record demos. Didn't go well. Claude asked for permission for every browser action, auto-edit on didn't help, but more importantly Claude didn't seem to be able to see things obviously on the page

Elon making a waifu factory was really not on my bingo card this year

Good tip for when Claude.md starts to get too large: break up your instructions into context-specific instructions stored in other .md files. e.g "When we work on the frontend, make sure you follow all the instructions I outlined in claude-docs/frontend.md"

OpenAI-model-drop day is the new no-meeting-day

GPT-5 release marketing feels immensely better planned than any of the other releases that happened in recent memory

"shower thoughts" are under siege there's so much content available in every form factor that it's easy to mute one's own thoughts. Quiet car rides, showers, puttering in the morning, staring in the distance on the subway etc.. are all petri dishes for one's own thoughts

you can probably purge bots entirely from this platform if you delete most accounts whose handles end in 5 seemingly random digits

Trying to answer the question “what is the state of AI tooling for engineers today” is funny because the answer literally changes every week

having poked around with the model, this makes more sense (in a way)

This is the best advice when dealing with the vibe-based hivemind