I'm opening @whirl_sh for early access Most AI agents still sit idle, waiting for a prompt. Not good enough for ops teams & messy real-world problems We need AI that runs in the background 24/7, proactively solving issues as they happen — with approval and visibility so it doesn't go fully rogue (🦞) At @packfleet I spent four years watching our ops team keep a dozen plates spinning across spreadsheets, Slack, email, WhatsApp, phone calls, and other tools that didn't talk to each other Whirl sits across your stack and proactively solves issues as they come in, leaving the team for the work that needs a human in the room (my good friends at Idylliq are already running their Paris property ops on it. Guest comms, booking admin, cleaners, owners, maintenance...) If your days are full of coordination work across a dozen tools, good chance Whirl can take a chunk off your plate. DMs open for early access or sign up on the site

LLMs are incredible at gathering all the right details, surfacing them in the right place, and even suggesting roughly what to write. Just don't ever make them write the final words

Would be great to eliminate reasoning levels as the next advancement in LLMs. Isn't it feasible to at least learn a router for that? It's nuts that Opus 5.5 low to medium are two totally different models but then you get almost nothing for another 5x in spend

We really need to talk about time-per-task for open models Here's similar aggregated runs on @whirl_sh: 5.6 Luna spends <1m per task to GLM-5.3 flash's 6m avg, for IMO very similar real world performance. From early tests MiMo v2.6 is a step in the wrong direction too

"Antigravity built in" - sounds pretty exciting if you don't know what it is

Isn't it about time this bar went straight to an AI assistant instead of web search?

I'm actually surprised Shopify didn't have a "search across all stores" before. I would've used that even in a pre-agent world

AI: "Short version: ..." Narrator: It was not the short version

"This is nothing new, it's just X combined with a Y! I could've done this easily!" "So why didn't you?"

Have you noticed how some Instinct messages sound templated on WhatsApp? "Time sensitive: [message] - reply here and I'll explain". Among many annoying things with WhatsApp for business is you can't send arbitrary messages to users if there haven't been any replies in the last 24 hours. I really hope Meta see the light and change this before we see the light and migrate to iMessage / RCS

So hard to see a winner takes all market for Instinct. 0 network effects, no UI and no contract / tie-in. So you just move to whoever has the slightly better service. (this could change if memory starts to matter but I highly doubt it, you can always extract memory)

Ever notice how AI never shows remorse or says "I'm sorry"? It's always "My bad" or "That was on me".

Fable consumes ~10% of my 5 hour quota on a Max (5x) plan in about 1.5 minutes. This is absolutely ridiculous

At some point I think I installed a PR review skill on Claude that launches like 10 sub-agents and obliterates my quota (or maybe my harness comes with it, I don't even know at this point, and too busy to fix it) Now every agent review prompt has to contain this at the end

Instinct keeps pinging me about stuff that looks superficially important but anyone with a tiny bit of real world context would know it doesn't need a direct ping. Biggest struggle in getting personal assistants to work well IMO

I'd like to propose a change of phrase from vibe-coding to doom-coding