I’ve been working on a small experimental gateway that sits between agents and customer-facing execution paths and forces decisions through policy + approval before anything irreversible happens.
v2 I just shipped adds:
• risk scoring on drafts • policy path traces • approval chain previews • highlighted spans showing what triggered the block • admin review endpoints
The motivation is exactly what people are pointing at here: once agents can transact, marketplaces end up banning them unless there’s a way to pause, inspect, and assign responsibility at execution time.
Curious what failure modes you’d want intercepted first if eBay or Amazon ever exposed agent purchase APIs.