Five decision gates for coding agents with Jev
The author walks through five gates in one iteration (intake, route, tool, verify, close), where the LLM only writes and Jev makes typed rulings via Choice and Score, plus three metrics to watch.
most coding agents still wake the writer to pick the fork workers ship the work. Jev owns the gate. the LLM only writes. one iteration, five gates: 1 → intake: one Noul before anything runs. in scope, or a person sees it first 2 → route: a Choice picks the worker (research, write, review). act at 0.82+. below that goes to review 3 → tool: allow / confirm / block. the LLM only writes the arguments 4 → verify: a Score on what comes back (drop, compress, keep). unsure answers go to a human 5 → close: Jev says done above 0.75. code then proves postmortem.md exists the LLM shows up in 2 of 5 gates. the rest is a typed answer, a check, or a person. don't rebuild the agent. hand Jev the one fork it hits most, usually the tool gate, and watch three numbers for a week: cost per finished job, wall-clock per finished job, how often escalation fired and whether it was right. if escalation never fires, the threshold is theatre. the agent is unsupervised with extra steps.
This post quotes another post: x.com/merccante/status/2102478458727113139