OpenAgentFlow Secures Heterogeneous AI Agent Fleets System-Wide

As AI agents evolve from isolated assistants to complex, heterogeneous systems where multiple agents, planners, and execution backends interact within a shared environment, ensuring safety becomes a critical system-level challenge. Existing safeguards often operate in silos, leading to fragmented enforcement, obscured risks in multi-step action flows, and limited support for auditing or evolving policies. OpenAgentFlow addresses these issues by proposing a novel control-plane/action-plane architecture. This system unifies all agent-generated actions—including GUI actions, API calls, tool calls, and LLM-generated invocations—into a single "AgentEvent" stream. Each event is then routed through a shared Policy Enforcement Point before being committed, allowing for consistent governance. The control plane maintains provenance, session state, audit records, and updatable policies, enabling new rules to take effect without modifying individual agents, prompts, models, or execution paths. Demonstrated on Android, OpenAgentFlow achieved high accuracy (94.0%) and attack block rates (95.3%) on an action-event benchmark, and successfully adapted to dynamic policy changes. These results confirm that OpenAgentFlow provides a practical, shared enforcement boundary for managing safety across diverse AI agent fleets.