Posts on X before Jun 25, 2026, 2:01:10 AM UTC

current agent work rule: the model is the least interesting part of the system. the loop around it matters more: repo read plan edit test log read retry handoff when that loop is visible, the work gets much less mystical.

vibe coding got popular in web apps because the browser is a cheap test harness. systems work needs harder proof: logs fixtures benchmarks trace spans failure repros rollback state give the agent those handles and it has something real to steer by.

if your agent setup needs babysitting, write that part down. where did it get stuck? which file did it read? which command failed? which check passed? what did you fix by hand? that is the map for making it less annoying tomorrow.

ok ok, glm 5.2 might be cracked they added 1m context, and with max reasoning, it's the first time an open model feels useful for what I use LLMs for which is exactly what matters for real work

ai can't automate you staring at your terminal deciding if you should refactor everything or just ship the hack or procrastinating on the agent that's supposed to automate your procrastination

I mean, can you imagine what things are going to be like a year from now? Hell, in six months, can you imagine the releases we're going to see before 2027 even starts?

You know, thanks to AI, I no longer ever have to touch my printer's settings. Even the latest printers have abominable settings. Honestly, anything via a menu feels so outdated when I can just tell my LLM exactly how I want it and it will script up whatever I need perfectly.

i know it sounds obvious, but maybe the most slept-on skill in agentic AI: context engineering BUT YOU shouldn't do the context gathering this is why the explosion of everyone using loops feels so sudden, but loops have been around or years regarding AI

the state of LLMs rn btw I went back a step, worded it differently, and it did it for me, go figure

the biggest mistake i made early on was ignoring tools that just wanted to do one thing well. no bloat, no dashboard, no integrations for the sake of it just clean input, clean output, reliable every time. whoever builds that kind of tool, builds something that lasts

what happens if the government just shuts down not only fable but opus and gpt 5.5 too has anyone actually pictured that or are we all just kind of assuming these models stick around forever + unlock further

10 tools I am using right now to build literally anything I want, June 2026 snapshot: 1. Pi (super custom) 2. Codex (CLI) 3. Codex (App) 4. Claude Code (CLI) 5. Blaxel 6. Obsidian 7. Linear 8. Slack 9. Cmux/Tmux 10. WisprFlow