Provenact runtime stops AI agents from using stale authorizations
New Provenact system prevents agents from acting on outdated approvals in finance and cloud workflows
AI agents are moving from advisory tools to systems that execute real-world actions like issuing refunds, reserving inventory, and initiating financial transfers. But existing safeguards often check permissions only at request time, missing situations where budgets, inventory, or approval states change before the action executes. The result is stale authorization: an agent acts on an approval that is no longer valid.
To solve this, researchers Yuxiang Peng and Xiaodi Wu propose Provenact, a runtime architecture that treats policies as reviewable programs while coordinating the state and effects needed to preserve their decisions. The core contribution is policy-state serializability, a correctness condition requiring every committed effect to be explainable as authorized against the policy state immediately before it occurs. In experiments with a PostgreSQL-backed prototype, Provenact prevented stale authorizations that baselines missed, preserved delayed approvals without blocking unrelated work, and avoided violations in a procurement workflow with shared budgets. The system also keeps most policy evolution in policy text rather than trusted provider code, making governance more auditable. Provenact offers a practical path for adding stateful governance boundaries to existing agent frameworks and provider-backed systems.
- Provenact defines 'policy-state serializability' to prevent stale authorizations in concurrent agent workflows
- PostgreSQL-backed prototype outperforms baselines that pass policy state as ordinary request context
- Policies remain reviewable programs, not opaque provider code, enabling easier auditing and evolution
Why It Matters
As agents handle financial and cloud operations, Provenact provides a correctness guarantee to prevent costly, unsafe actions.