Leaked Anthropic schematic: 8 agents on Jev + Opus 5.5 close tasks in 3 iterations
A circulated 14-block architecture diagram assigns planning, research, coding, execution and review to separate agents, with Jev holding decision memory across them; it claims an 87% task closure rate versus 20% for simple prompting.
THIS FILE WAS BURIED AT ANTHROPIC FOR HALF A YEAR, IN IT 8 AGENTS ON JEV + OPUS 5.5 CLOSE A TASK IN 3 ITERATIONS i hunted for this document for weeks, it sits in no public doc inside is one schematic with 14 blocks, and it is worth more than any $2000 agent course Jev cuts an open-ended task into small verifiable decisions, Opus 5.5 supplies the reasoning and the code under them the planner splits the goal into steps, the researcher reads the repo, the coder writes, the executor runs the commands, the critic hunts for holes Jev memory holds context across all of them, so the eighth agent knows what the first one decided and no step gets rewritten twice every step is a typed decision with a small structured output instead of a wall of text that is why Jev + Opus 5.5 lands the result in 3 iterations where a plain prompt does not get there in 10 the gap over distance: 87% of tasks closed against 20% on simple prompting one run through this schematic costs cents and closes a ticket an agency bills $4000 for what comes out is not just code, it is code + tests + docs + confidence score + rollback plan
This post quotes another post: x.com/noisyb0y1/status/2102309802072272956