The rise of Large Language Models (LLMs) has revolutionized how we think about automation and collaboration. These powerful AI systems can now act as autonomous agents, capable of reasoning, decision-making, and working together in multi-agent systems. But what happens when these AI agents need to collaborate with humans in complex workflows? Current process modeling languages, including the widely-used BPMN (Business Process Model and Notation), fall short in capturing this new reality.
In our paper “Towards Modeling Human-Agentic Collaborative Workflows: A BPMN Extension”, accepted at the 51st Euromicro Conference Series on Software Engineering and Advanced Applications (SEAA’25), we address this gap by extending BPMN to support human-agentic collaborative workflows. This work provides a structured approach to model the complex interactions between humans and AI agents in business processes.
Modeling Human-Agent Collaboration
Traditional process modeling languages, such as BPMN, were designed for human-to-human collaboration. While BPMN excels at representing workflows involving human participants, it lacks the specific constructs needed to model agents beyond simple automatisms, in particular:
- Decision-making processes: How are decisions made when multiple agents propose different solutions?
- Collaboration patterns: How do we formalize the different collaboration types between agents?
- Agent behavior and reliability: How do we represent the non-deterministic nature of AI agents and their varying levels of trustworthiness?
- Reflection strategies: How do agents improve their outputs through self-reflection, cross-reflection with other agents, or feedback from humans?
To better understand the current context, consider a simple bug resolution process in software development. A user reports a bug, an AI reviewer agent validates it (using self-reflection to double-check), two coding agents independently propose solutions, and a human maintainer makes the final decision. This seemingly straightforward process involves complex agent interactions that standard BPMN cannot adequately represent (see the paper for full details).
Our Approach: A BPMN Extension for Human-Agentic Workflows
We developed an extension to BPMN that introduces new modeling constructs specifically designed for human-agentic collaborative workflows. Our extension addresses the following areas: (1) agent profiling, (2) agent reflection, and (3) agent collaboration. We modeled the requirements for each extension point, based on the state of the art.
BPMN treats all participants as generic actors, but agents have distinct properties that need explicit modeling. Our extension introduces mechanisms to identify lanes and pools as agents, while enabling the specification of agent roles and their varying levels of trustworthiness. This agent profiling becomes crucial when modeling workflows where decisions depend on the reliability of different participants. Furthermore, to refine their answers, agents can make use of reflection strategies to review their actions by inspecting themselves, other agents, or even human participants. We provided the adequate syntax to represent the different kinds of reflection.
Regarding the collaboration between (or with) agents, we modeled the different possibilities under cooperation (voting-, role- or debate-based) and competition. We extended the gateways’ syntax to be able to model these collaboration scenarios within pools, and the message flows for inter-pool interactions. On the other hand, we also provided support to represent the different ways we can merge the agents’ output. From the literature, we represented different merging strategies (e.g., leader-driven, composite, by voting majority).
The following figure, illustrates the example presented before, using our notation:
The different agents are denoted with agentic lanes, and have a trust score attached (see three top lanes). The second task (see Check bug validity) is an agentic task that applies self-reflection to the output. To represent the collaboration between agents, we use the agentic gateway, since agents are represented as lanes of the same pool. The collaboration strategy applied is role cooperation, as denoted by the notation. After stating the collaboration strategy, the flow is divided towards the agents that collaborate. All flows from this collaboration are merged into an agentic gateway following the leader-driven strategy, as denoted by the notation. The remaining elements are compliant to standard BPMN.
Tool support
We have implemented our extension in an open-source modeling editor tool available on GitHub (check here). The tool includes a main view of the diagram and a palette to allow the user to drag and drop the elements into the diagram. As a proof of concept, we have provided a few examples to illustrate how to represent different scenarios (check them in “Further Examples” here).
Conclusion
The integration of AI agents into existing workflows is not a distant future, it’s happening now. Our BPMN extension provides practitioners with the tools they need to model these complex human-agentic collaborative workflows. By building on established standards and providing concrete tool support, we aim to make this transition as smooth as possible for organizations embracing AI-augmented business processes. The next steps we will focus on are
developing more sophisticated governance mechanisms, uncertainty propagation models, and code generators that can automatically create executable workflows from our extended BPMN models.
We look forward to presenting our findings at SEAA’25 and exploring further applications of our methodology in real-world scenarios. The future of work is collaborative—between humans and AI—and we need the right modeling tools to navigate this new landscape effectively!
PhD student at the SnT-University of Luxembourg. Currently working on providing modeling methodologies to the definition of LLM-agents systems, such as domain-specific languages.
Impressive.
Was it all done simply by using ECORE and Aconite? Or was there any need for further Sirius adaptation?
I am curious how effective Aconite is for automating the creation of domain specific workbenches.
Hi, it was all done with Aconite and ECORE.
Cool. Thanks. Seems promising.