Intervention is deliberate runtime control that lets authorized parties or systems interrupt, restrict, or redirect automated behavior. It matters when legacy automation repeats a failed action, such as retrying a blocked payment thousands of times or continuously calling an unavailable service. Rate limits constrain repetition; kill switches can block new actions, revoke tool access, cancel safe in-flight work, or isolate a workflow. Completed actions require separate compensating or recovery actions. section 1.3
A credible intervention boundary transfers control safely when confidence narrows, constraints fail, or conditions change. The design must specify what triggers transfer, where control goes, what context and authority move with it, and what happens if escalation fails or is delayed. section 2.1.5
Operational intervention requires explicit decisions about:
- Authority boundaries and policy checks that determine when autonomous action remains legitimate. section 2.2
- Thresholds, escalation triggers, rate limits, hard stops, and observable decision boundaries. section 1.3 section 2.2
- Named ownership for escalation, intervention, and ongoing stewardship of delegated authority. section 4.6
Containing a runaway payment workflow
When a workflow repeatedly retries a blocked payment, a rate limit constrains further attempts. A broader kill switch can block new actions, revoke tool access, cancel safe in-flight work, or isolate the workflow. Any payment already completed needs a separately designed compensating or recovery action. section 1.3
Intervention makes bounded autonomy governable rather than merely permissive. Architects define the supervised decision space, while operations handle runtime events; accountability remains with the named business owner, and authority must be reviewed so it does not silently expand. section 2.2 section 4.6


