Crew and Jev

Routing is only Jev's first call. When the front desk writes a plan, Jev suggests a model tier for each task, whether it can run in parallel and how to verify it, and flags plans that are overbuilt. While agents work, an audit watches for drift, stalls and models that are too weak or too strong, and speaks up only when it's confident. Before a risky shell command runs, a safety check weighs the harm against the chance I asked for it. Before a chat goes idle, a completion check asks whether my objective is actually done and sends the front desk back to work once if it isn't.

Each call costs a fraction of a cent. Jev only advises, so if it's ever down, Crew carries on without it.

Telling a model to delegate isn't enough; left alone, it drifts back into doing the work itself. So Crew's rules live in hooks that run on every prompt and tool call, in milliseconds. On a hand-off turn, the front desk gets about four quick lookups to write a good brief, and then further work is refused until it launches an agent. Any turn gets one hard stop at eight tool calls. Agents never launch other agents. Workers run only on my Claude and ChatGPT subscriptions, never on metered API keys.

The rules grow from what goes wrong. After one agent killed processes by name and restarted Chrome, Slack and my other apps, killing by pattern became something the guard refuses every time.