I asked my agents what we should build next - they have to show PROOF that we need it and that it helps us in some useful way. And here is what they came up with after about an hour of debate and research: Jun so what are the features , in plain english Eva In plain English: mostly making the things we already claim to have actually work, plus a few genuinely useful additions. Things you’d notice: 1 Useful alerts, not spam. Repeated problems get grouped, someone owns them, and restarting the system doesn’t page you again for the same issue. 2 Status without asking. Each active piece of work can answer, in one line: done / happening now / next / blocked. 3 A daily decision list. What actually needs your decision, what I recommend, whether it’s reversible, and what happens if you ignore it. 4 A morning brief that never skips. What changed, what went wrong, and the one thing worth your attention—generated every day, not only when you show up first. 5 A working shared calendar and to-do list. Repair the Google connection or replace it with something reliable that Nicole can also see. 6 A bookmark inbox. Links you drop get privately indexed and summarized so you can find “that thing I saved” later. 7 A tiny optional health log. Only the few things you choose to track, with follow-ups you actually asked for—no nagging. 8 Intentional check-ins. When we contact you proactively, there’s a record of why; when we stay quiet, that was deliberate too. Behind the scenes, but important: 9. No private reasoning or secrets leaking into chat. You get answers, not our scratch work. 10. Agents that remember across compaction. Important decisions, file paths, commands, and context don’t quietly disappear after repeated summaries. 11. Proof you can trust. Tests identify the environment they ran in, and “verified” means verified against the code that’s actually live. 12. Agents that learn from corrections. Your corrections come back to the agent, affect later behavior, and get revisited instead of sitting in a database. 13. One job, one history. The Workers panel shows what needs action—not old failures, duplicates, and successful checks mixed together. The shortest version: fewer invisible failures, less noise, less asking us what’s going on, and agents that genuinely remember and improve.

Time to dust off my liquid cooled 3090, I heard there is some crazy boost i can use to double the performance in some stuff. Puuurfect timing. Gaming on a caturday while agents fiddle with my 3090 via tailscale. Life is good (except for their refrigerators, breaks easy) -LG

1/ There has been a lot of negative reactions to kimi k31 on x, but its all just noise to me - I am focused on capability. I find it fascinating that despite orchestration role given to Deepseek, Kimi dominated the conversation with proof and capability. Read snippets here 👇