Towards Modeling Human-Agentic Collaborative Workflows: A BPMN Extension

Adem Ait, Javier Luis Cánovas Izquierdo, Jordi Cabot

This paper, recently published by researchers from the University of Luxembourg and UOC, addresses a burgeoning challenge in the age of AI: how do we formally design and visualize workflows where humans and autonomous AI agents work side-by-side?

TL;DR

The authors propose a formal extension to the Business Process Model and Notation (BPMN) standard specifically designed for "human-agentic" workflows. By introducing new primitives for agent trust, reflection strategies, and collaboration patterns, they provide a structured way to model complex interactions between humans and LLM-based agents that standard business process tools cannot currently express.

What Problem Does This Paper Address?

As Large Language Models (LLMs) evolve into autonomous agents, they are increasingly being integrated into multi-agent systems (MAS) to solve complex tasks. However, these agents don't just work with each other; they must collaborate with humans.

Current workflow modeling standards like BPMN fall short in this "mixed" environment. In standard BPMN, an AI agent is typically treated as just another "Service Task" or a "User." This masks critical nuances of agentic behavior, such as:

  • Non-determinism: How confident is the agent in its output?
  • Reflection: Does the agent double-check its own work (self-reflection) or ask a human for feedback?
  • Conflict Resolution: If two agents propose different solutions, how is the final decision made (voting, debate, or a manager's call)?

Without a formal way to model these, developers rely on ambiguous text annotations, leading to "spaghetti" diagrams that are hard to read and impossible to automate reliably.

Methods

The authors utilized the built-in extension mechanism of BPMN 2.0 to add "agent-aware" features without breaking the existing standard. Their approach covers three main areas:

  1. Agent Profiling (AgenticLane): They extended the concept of a "Lane" to include a Profile (defining roles like 'manager' or 'worker') and Uncertainty (a trust score from 0-100).
  2. Agent Reflection (AgenticTask): They introduced a specific task type that can be assigned a reflection mode: Self-reflection, Cross-reflection (feedback from other agents), or Human-reflection.
  3. Collaboration & Merging: They extended Gateways and Message Flows to define how agents cooperate. This includes specific logic for Voting-based, Role-based, Debate-based, and Competition-based collaboration.

To make this practical, they developed a graphical notation using a consistent "agent marker" (a small robot-head icon) with letter modifiers (e.g., 's' for self-reflection, 'v' for voting) to keep the diagrams clean but informative.

Datasets

Not discussed in the paper. As a methodology and modeling paper, it focuses on the architectural framework and notation rather than training on a specific dataset.

Key Insights and Ideas

  • Trust as a First-Class Citizen: By attaching a "trust score" to an agent's lane or task, the workflow can programmatically branch. For example, if an agent's confidence falls below 70%, the workflow can automatically trigger a human review.
  • Formalizing the "Inner Loop": Reflection is a core part of how LLM agents improve. Moving this from a messy loop of arrows in a diagram to a single property of an AgenticTask significantly reduces visual clutter.
  • The "Manager" Role: The extension introduces a "manager" role for agents. In a role-based collaboration, the "manager" agent is the one responsible for selecting the valid output from several "worker" agents, mirroring human organizational structures.

Key Results

  • Expressiveness: The authors demonstrated that their extension could model a complex bug-resolution process (involving a human user, a human maintainer, and three distinct AI agents) much more clearly than standard BPMN.
  • Proof of Concept: They implemented a functional modeling editor using the Sirius framework (an Eclipse project). This tool allows developers to drag and drop these new agentic elements into their process diagrams.
  • Notation Clarity: The proposed notation follows "semiotic clarity" principles, ensuring that the new agentic symbols are easily distinguishable from standard human tasks while remaining familiar to BPMN users.

Limitations and Future Work

The paper is an exploratory first step. The authors acknowledge several areas for growth:

  • Lack of Execution: Currently, the tool is for modeling only. The next step is to build code generators that can turn these diagrams into executable code for agent frameworks like LangGraph or CrewAI.
  • Uncertainty Propagation: While they introduced trust scores, they still need to define how that uncertainty "flows" through a process (e.g., if a low-confidence agent provides data to a second agent, how does that affect the second agent's confidence?).
  • Empirical Validation: The authors plan to conduct user studies to see if these new notations actually help engineers design better systems compared to traditional methods.

Summary updated

Abstract

Large Language Models (LLMs) have facilitated the definition of autonomous intelligent agents. Such agents have already demonstrated their potential in solving complex tasks in different domains. And they can further increase their performance when collaborating with other agents in a multi-agent system. However, the orchestration and coordination of these agents is still challenging, especially when they need to interact with humans as part of human-agentic collaborative workflows. These kinds of workflows need to be precisely specified so that it is clear whose responsible for each task, what strategies agents can follow to complete individual tasks or how decisions will be taken when different alternatives are proposed, among others. Current business process modeling languages fall short when it comes to specifying these new mixed collaborative scenarios. In this exploratory paper, we extend a well-known process modeling language (i.e., BPMN) to enable the definition of this new type of workflow. Our extension covers both the formalization of the new metamodeling concepts required and the proposal of a BPMN-like graphical notation to facilitate the definition of these workflows. Our extension has been implemented and is available as an open-source human-agentic workflow modeling editor on GitHub.
BibTeX
@inproceedings{ait2025towards,
  author = {Adem Ait and Javier Luis Cánovas Izquierdo and Jordi Cabot},
  title = {{Towards Modeling Human-Agentic Collaborative Workflows: A BPMN Extension}},
  booktitle = {SEAA},
  year = {2025},
  doi = {10.1007/978-3-032-04190-6\_22},
}

Loading related content...

Figure