Agent architecture that uses Jev as the decision layer to save Claude quota
A division-of-labor setup where Claude only writes text, Jev handles routing, filtering, verification and risk checks, and code executes actions, claimed to cost $0.42 per 10,000 decisions.
What they stay silent about so you keep burning through your Claude limits > JEV savings and automation The ideal plan: Claude only writes. Jev makes decisions, code executes actions, and every step is correct Where to start: AGENTS.md > when to use Jev, when not to use it state/build_state.py > goal, task, workers, completed, missing, constraint, evidence, pending Jev decides (questions/) http://route.py > selects the model power level for the task next_worker.py > passes the baton to the next worker http://relevance.py > filters out garbage: selects only useful tool output http://done.py > verifies the real result, not just the execution status http://risky.py > blocks dangerous actions (payments, data transmission, etc) Code acts (rules/) hard_rules.py > a 10-step limit and a strict ban on publishing raw content. thresholds.yaml > execution only upon high confidence; for critical/risky operations, the threshold is 0.95 Claude writes (workers/) http://research.py > source collection and working notes. http://writer.py > preparation of drafts and briefs. The only zone in the system where text generation happens Proof receipts/decisions.jsonl > decision log: candidates, selected ID, validation results, and fallback scenario. evals/ > a dataset of dozens of personal traces used to calibrate all thresholds Guard (hooks/) pre_tool_use.sh > validation of each command strictly before execution. http://stop.sh > final checklist: verifies readiness criteria before stopping the agent 10,000 decisions for just $0.42 at an average speed of 300 ms Division of labor: LLM creates content, Jev makes decisions, code executes Using this is a genuinely smart move. Save it so you don't lose it
This post quotes another post: x.com/bober_smart/status/2078784709253841039