Beyond Rule-Based Workflows: An Information-Flow-Orchestrated Multi-Agents Paradigm via Agent-to-Agent Communication from CORAL
arXiv preprint arXiv:2601.09883, January 2026
Rule-based workflows are the standard way to make multi-agent systems reliable: fix the graph of who calls whom, and behavior becomes predictable. The cost is that the system can only handle tasks whose shape someone anticipated in advance, and real long-horizon tasks rarely cooperate.
This paper drops the fixed workflow entirely. Agents communicate in natural language over an A2A protocol, and a centralized information-flow orchestrator monitors task state and adapts coordination dynamically instead of executing a predetermined graph.
Results on GAIA:
- 63.64% pass@1, versus the workflow-based OWL baseline.
- +8.49% improvement at comparable token usage; the gain is not bought with extra inference budget.
- Notably better behavior on edge cases and long-horizon task decomposition, which is exactly where fixed workflows fail.
Read together with Anemoi, the two papers bracket a design question I find genuinely open: how much central coordination does a multi-agent system actually need? Anemoi removes the central planner; this work keeps a coordinator but strips the fixed workflow. Both beat the rigid baseline, which suggests the rigidity, not the centralization, was the problem.