Researchers pair Claude Opus 5.5 with Jev in a layered agent architecture
A researcher describes an architecture where Opus 5.5 plans the work and Jev answers each branch as a typed call, claiming 7x lower model spend, zero-token routing decisions and no reward hacking.
this is pure f*cking treasure a team of researchers found how to orchestrate Claude Opus 5.5 and Jev together, and it completely breaks the agent scaffolding wall most developers try to make an agent reliable by routing every decision back through the model. you burn tokens on choices that have an exact rule and freeze your routing logic into a prompt that drifts within 5 edits this architecture eliminates human-written routing prompts and splits the cognitive stack: search director: Claude Opus 5.5 plans the work, querying a persistent decision graph where routing and scoring reduce to single typed calls typed router: Jev answers every fork in code, returning an option with a probability instead of prose zero context bleeding: a routed fork never enters the chat context, so a thousand decisions cost the same context as zero evaluator isolation: candidate code never touches the scoring process, stopping reward hacking and test leakage cold handoff contract: every step passes task, inputs, done-condition and confidence as a file, so a crashed session loses nothing the metrics outclass single-context agent loops: 7x lower model spend than running the same loop through one context window 0 tokens generated on routed decisions sub-500ms on forks that used to cost a full frontier round trip 0% reward hacks: candidate-controlled metric tampering eliminated Opus directs the plan. Jev picks the branch you stop writing rigid routing prompts. you let the type system own the decision loop I'll leave some ready-to-copy Jev builds in the article below ↓ ↓
This post quotes another post: x.com/0x_rody/status/2103165281149354256