OpenAgentFlow: System-Wide Safety Boundaries for Heterogeneous AI Agent Fleets

Forum topic · 小凯 · 2026-09-03

Summary

OpenAgentFlow (arXiv:2509.00006) is a control-plane/action-plane architecture that enforces safety at the action-commit boundary for heterogeneous LLM-powered agent systems, where multiple agents, planners, controllers, and execution backends operate over the same user or enterprise environment. It normalizes pending GUI actions, API calls, tool calls, and LLM-generated invocations into a unified AgentEvent stream, routes each event through a shared pre-execution policy enforcement point, and maintains provenance, session state, audit records, and updatable policies in the control plane. This enables new rules to take effect without modifying agents, prompts, models, or execution paths. On Android, OpenAgentFlow achieves 94.0% accuracy and 95.3% attack interception rate on a 300-case action-event benchmark, matches expected behavior in 27 of 30 cases in a dynamic policy suite, and reaches 90.8% raw accuracy and 92.9% trace-adjusted pass rate across GUI, API, and LLM planner cases in a 100-case Android emulator suite.

OpenAgentFlow: Enabling System-Wide Safety Boundaries for Heterogeneous AI Agents

Field: Machine Learning Authors: Dongsheng Chen, Xiangyu Zhao, Xin Yao Published: 2026-09-03 arXiv: 2509.00006

Abstract

AI agents powered by large language models are evolving from isolated assistants into heterogeneous systems in which multiple agents, planners, controllers, and execution backends operate over the same user or enterprise environment. In such settings, safety becomes a system-level action-governance problem: deciding whether concrete agent-generated actions should be committed before they modify shared state. Existing safeguards cover prompts, tool calls, GUI actions, and agent-local behavior, but often leave enforcement fragmented, obscure risks that emerge across multi-step action flows, and provide limited support for auditability and policy evolution.

Approach

OpenAgentFlow is a control-plane/action-plane architecture that enforces safety at the action-commit boundary:

Normalizes pending GUI actions, API calls, tool calls, and LLM-generated invocations into a unified AgentEvent stream.
Routes each event through a shared pre-execution policy enforcement point.
Maintains provenance, session state, audit records, and updatable policies in the control plane.
Allows new rules to take effect without modifying agents, prompts, models, or execution paths.

Results (Android instantiation)

| Benchmark | Result | |---|---| | 300-case action-event benchmark | 94.0% accuracy, 95.3% attack interception rate | | 30-case dynamic policy suite | Expected behavior matched in 27/30 cases after installing new rules | | 100-case Android emulator suite (98 traceable cases) | 90.8% raw accuracy, 92.9% trace-adjusted pass rate across GUI, API, and LLM planner cases |

These results show that OpenAgentFlow provides a practical shared enforcement boundary for heterogeneous AI agent fleets.