In the original language. Each post links to the conversation.
12 of 85 posts
▶ Watch videoMost human-in-the-loop systems put the human at the end to review the output. In my agent workflows, the useful approval happens before execution: plan, permissions and stop conditions. The agent can then work inside clear boundaries. This clip from @gregisenberg captures the
Open on XI never understood what was so cool about token maxing. I even looked at the meaning and surprisingly it was really just spend as much as you can. This didn't make sense to me at that time. Now it seems people are (and will be) thinking more about token efficiency!
Open on XMy mother in law came home to spend the week. When she was at the living room she asked my wife: "Is Ricardo ok? I heard him talking to his computer all day" Yeah, that's me vibe coding!
Open on XThis is insane
Open on XEveryone's hyped about agent loops since @steipete post and @bcherny . The part that got me isn't running agents solo. It's that the loop only works when your judgment is already baked into it. Distribution still decide who wins.
Open on XOwn your memory. The AI race is so crazy that everyday there's a change in the best model or harness. If you own your memory you can take your data anywhere easily.
Open on X
The more I work with AI agents, the more parallels I find with how humans work. Memory is one of them. My agent now has a sleep cycle: light sleep, REM, deep sleep. Engineering has been imitating biology in more places than it seems.
Open on X
Lost months of context migrating between AI tools. Decisions, planning, brainstorms from entire projects, all stuck in the old service. The fix wasn't picking a better AI. It was owning my memory. The AI you use matters less and the memory you feed it matters more.
Open on XOpus took the direction, picked a path, made a reasonable call when context was missing, came back with the work done. Like the teammate you brief once and they figure out the rest.
Open on XRan the same autonomous agent on Claude Opus 4.6 and GPT-5.4. Same config. Same task. One shipped. The other stalled. The difference is something benchmarks don't measure.
Open on XMost companies are using AI to produce more stuff, not to make better decisions. That’s why a lot of “AI adoption” feels fake. More output, same confusion. The real unlock is not content velocity. It’s reducing friction in research, decision-making, and execution.
Open on X@CenturyLink I'm really tired of your service! The internet is never stable I need to work! You say it's fixed but next day we have the same problem again Who live in Fort Myers don't use their service!
Open on XFor ideas that need a little more room.
Read the articles