Stalls, failures, runaway loops, no-shows. Regisseur watches for them, proposes a way through, and escalates, and a person always authorizes the recovery. Self-healing operations, not autonomous ones.
A platform-level coordinator monitors case progress and surfaces work that's stuck, and manages the cross-party inquiries (chasing the outside party a case is waiting on) so a case doesn't quietly park forever.
When a step fails, the recovery engine identifies the blocked downstream work, finds candidate routes around the failure (bypass or parallel unblocked paths), and scores each by timeline impact, days delayed, SLAs at risk.
It proposes; the process manager chooses. Recovery is never silent.
Two safety levels. The workspace-wide emergency brake halts every agent, queue, and outbound message. Below it, a per-job circuit breaker trips a single runaway case (guarding against execute/cascade loops and runaway cost) without stopping the workspace.
Every step carries an SLA clock. Escalation policies define who gets notified, on which channel, and when a budget is at risk. SLA projection flags at-risk cases before they breach.
Bring a process that stalls in real life. We'll show the coordinator surfacing it, the recovery engine proposing scored routes around the failure, and a person authorizing the one that ships.