JEV, Opus 5.5 and GPT Dots run a task loop on one screen
Tasks are split between a planning brain, Opus 5.5 as a spider messenger and four worker agents, showing about five closed loops per minute on one screen, with the spider carrying three tasks at a time and rarely running empty.
HOLY SH*T! JEV + Claude Opus 5.5 + GPT Dots in one closed loop, live on a single screen. the spider carries 3 things at a time and almost never walks empty. most AI stacks send every task through the same pipe: something comes in, the biggest model looks at it, and the biggest model answers. it's expensive, slow, and most of the time it's overkill. i split it into a brain, a spider and four workers instead. here's the breakdown: 1. the brain: JEV - a 5,200-node mesh with 12 regions and 420 fibre tracts. it doesn't do the work, it plans it. every task lights up as a beacon on the cortex, right where it came from. → lesson: the planner's job is to decide what's worth doing, not to do it. 2. the spider: Opus 5.5 - it doesn't watch everything. it walks to the hottest beacons and picks up to 3 tasks. then it drops them at the workers and carries each result back to the spot on the brain it came from. → lesson: the expensive model should route and check, not grind. 3. the Dots agents: Echo, Sprout, Pocket, Stella - each has its own queue, bench and outbox. they keep working while finished results wait for pickup. when they're idle they run maintenance instead of sitting there. → lesson: workers shouldn't block on the courier. decouple the handoff. 4. you - you read one dashboard. you can trace every task from beacon to bench and back into the brain. the math: 1 planner → 3-slot courier → 4 workers → ~5 closed loops a minute → 1 screen. i'm breaking down one AI stack like this every week. if your team still routes every task through the biggest model, show this to whoever pays the API bill. would you let a cheap planner tell your most expensive model where to walk?
This post quotes another post: x.com/starmexxx/status/2108159970017488953