Jev in the control plane: an auditable agent setup in 8 steps
The author moves the decision layer off the LLM and onto Jev, adding Kafka, a toll booth and a human quarantine zone to run the first agent setup they didn't have to babysit.
Jev in the control plane is the first agent setup I have run without sitting behind it 99% of people put an LLM on every decision and pay 400x for an answer that takes 3 seconds. 1% run the decision layer separately. the whole shape is one line: prompt → orchestrator → Jev decision + p → toll booth → fast lane or human → tool → observations back 8 steps and the architecture is done: step 1 → log the action before you execute it, in the same DB transaction as the request. if the process dies, the action is still recorded. step 2 → put it on kafka partitioned by session_id. one agent session stays ordered, and you get replication and idempotency keys for free. step 3 → make the orchestrator stateless. plan, react, manage tool calls, track state. state lives in kafka and the execution log, never in the process. step 4 → send every decision to Jev, not to the LLM. route, classify, score, extract, branch. typed output, sampled in one parallel pass, 70 to 500 ms, $0.042 per million in and output free. call the LLM only when the answer genuinely needs strings. step 5 → build the toll booth: policy → authorisation → validation → idempotency. nothing reaches a tool without passing all four. this box needs 10K decisions a second, which is exactly why it cannot run on a model that takes 3 to 329 seconds. step 6 → use the calibrated probability as the gate. over the gate it runs in the fast lane. under the gate it goes to quarantine, a human sees the proposed action and the reasoning, and it is revalidated in full before it executes. approval never permanently bypasses policy. step 7 → never call external tools synchronously, and never blindly retry an ambiguous state changing action. idempotent outbox, then reconcile: safe → finalise, blocked → quarantine, unknown → human review. step 8 → run it in shadow for 20 to 50 real decisions. the agent works normally while every Jev route is logged next to what you would have done. promote only the routes that hold. you are done. the model was never the bottleneck. the decision layer was.
This post quotes another post: x.com/n01ennn/status/2096962591125905888