Stopped treating ALLOW as the end of the story. An ALLOW decision without execution tracking is just a suggestion. The agent could execute twice, or not at all, and nobody would know until something broke in production.
Started shaping a lifecycle that binds decision to execution to outcome. The flow is: evaluate, then optionally route through approval, then execute, then finalize. Every stage has a defined state and every transition gets recorded. No gaps.
Early thinking on claim/lease semantics this week too. The idea is TTL-based: an agent claims a decision, gets a window to execute, and must finalize before the lease expires. If it does not finalize, the system knows something went wrong. No silent failures. No zombie executions that nobody notices until the monthly invoice shows up.
Exactly-once execution becomes a core invariant from this point forward. Everything else is negotiable but this is not.