porsche hasn't said what engine is in the mission s, so i simulated what it might sound like: flat-8, 8000rpm, idle matched to this clip. no samples, every sound is a cylinder firing into a pipe insanely fun to play with :D

this becomes feasible when: - your agent can run the full stack locally and use the app itself to prove a change works - it has a good framework for reasoning about what it does and doesn't need your opinion for. both for product and engineering decisions - your agent can watch logs, db, replays after deploying to monitor the change and create followups/reverts if needed - your agent runs a good code review - you personally trust the model - your agents talk to a coordinator that manages your attention and prioritises decisions for you we have that with rook and opus 5.5

luna speed-pictionary i cant draw giraffes

here's a nice ui touch - the hovered sidebar item never moves while your mouse is over it. even as other items in the sidebar shift and move. so you never accidentally archive the wrong item because something shifted as you clicked.

jev is insane: live intent tracking as you type we've never before had access to judgement at this latency+cost level

fine-tuned model that only writes 72-char lines coworkers are not ready for my perfectly-justified messages

you can just build your own instinct agent. over imessage, answers in seconds, watches things for you around the clock. chuck this tweet at fable and it'll get it done in an afternoon. then you can worry less about what's happening with your email data. every piece exists off the shelf: - sendblue: imessage api. typing indicators, tapbacks, read receipts - cloudflare agents: the harness. one durable object = your assistant, alive forever, wakes in ms - cf sandbox: code execution - cf artifacts: durable storage that speaks git, memory you can read and diff - cf dynamic workers: triggers - cf agent memory: culls chat history into actual memory - browserbase: browser with real fingerprints for logged-in stuff run opus 4.8 on low thinking. ~2 second replies. the speed is what makes instinct feel good. pay for fast mode if you want. triggers, "when x happens, do y", three levels: 1. a dynamic worker polls the narrowest endpoint that changes. deterministic diff, did it change or not. no tokens, costs nothing. 2. on change only: one cheap llm eval. does this match what i asked for? 3. on match only: wake the main agent, it decides whether you'd actually want a text. cull old conversations periodically into curated markdown files, and keep the verbatim history somewhere the agent can search. build context as just [memory files + recent messages]. compaction and long-term memory, same system. never let the main agent do work. make it talk and delegate. subagents are trivial on cf agents: spawn another durable object with a brief and a callback, hand it a sandbox if it needs compute. gmail, easy for a personal app. spin up a gcp project, enable the gmail api. now it triages your inbox and drafts replies you approve. and the only thing reading your mail is a worker 'you' wrote, holding a token you can revoke. anyway, mine is currently looking for used surfboards and notifying me when it finds something good

fable: bake 2–3 days watching A2 warn + D1 canary hold zero, then D2 with the safety-net switch folded in. Every gated increment already baked clean and shipped (B2 verdict 2026-07-23, C1a removed at the C1 flip, D2 shipped on its pre-ship evidence). Canaries quiet: 0 severed legs. me: 🥺

slack's new agents api is pretty great

go to sleep, wake up to a 40k pr, throw a 🤘 and hit merge

second gh actions outage of the day is a sign to go touch grass

my multi-repo cloud agent getting to done-done

this is perhaps my favourite game. play it with zero context, you'll laugh, you'll cry, it'll change your life

building on the agents sdk has been an absolute pleasure (even when my state hits the 2mb limit)