New Blueprint Makes AI Agents Easier to Manage and Trust
This could make AI at work more reliable, saving time and reducing errors.
LLM-based agents can reason, use tools, and collaborate, but coordinating increasingly complex agentic applications is difficult to understand, monitor, and improve because decisions distributed across prompts, tool calls, and application logic make workflows implicit. The paper introduces Recursive Capability Composition, enabling agents to execute business processes and expose them as reusable Agentic Tools.
It presents a BPMN-based approach to Agentic Workflows that combines explicit orchestration with autonomous reasoning: SemTalk BPMN models are transformed into executable workflows using a LangGraph/Pregel runtime. Agentic Tasks introduce LLM reasoning at selected activities, while conventional BPMN controls process flows, deterministic decisions, system interactions, and human approvals.
Agentic Participants provide reusable interfaces for agent behavior or other executable processes. BPMN orchestration makes multi-agent systems more understandable and manageable, and explicit process models give developers and operators a shared view of workflow structures and execution progress, facilitating debugging, monitoring, and optimization. Enterprise agents also need organization-specific context: Process Models and Knowledge Graphs describe activities, participants, information objects, decisions, and relationships between processes, providing structured knowledge of how an organization operates to support informed agent decisions and actions. The contributions are model-driven execution, recursive capability composition, and enterprise context.
- AI agents can now follow visual flowcharts, making their actions clear and easy to monitor.
- These flowcharts can be reused as building blocks, saving time when creating new AI workflows.
- The system uses company knowledge to help AI make better decisions, but requires upfront planning.
Why It Matters
This makes AI at work more transparent and reliable, reducing errors and saving time.