A new arXiv paper, “KedgeFlow: A Programmable Control Plane for Enterprise AI Agents,” argues that enterprise AI agents can fail in specific ways when approval and resource execution are not tightly coupled. The authors describe cases where an agent executes an altered version of an approved action, reuses the same permission concurrently, or retries an operation whose response was lost.
The common cause, the abstract suggests, is that approval and resource execution lack a shared context. KedgeFlow is proposed as a programmable control plane to address this gap, though the abstract does not provide implementation details.
Because the paper is only represented by its abstract here, the full mechanism and evaluation are not available. The value of the work at this stage is in naming these failure modes clearly and pointing toward a control-plane approach for enterprise AI safety.