Giving AI agents a workflow map instead of the whole steering wheel

A visible plan for partly autonomous work

The preprint describes an approach implemented in SemTalk, a business-process modeling environment. It uses Business Process Model and Notation, or BPMN: a standardized way to draw activities, decisions, parallel branches and the participants responsible for work.

The proposal separates two questions. The process model determines when an activity happens and what comes next. An AI agent determines how to carry out an activity assigned to it. Conventional software and human tasks remain part of the same process rather than being replaced by autonomous reasoning.

That distinction matters because not every step needs a language model. A fixed routing rule can remain deterministic, while interpreting an ambiguous request can be delegated to an agent. Human interaction can occur at designated points.

Turning diagrams into executable workflows

SemTalk transforms BPMN models into executable graphs using a LangGraph/Pregel runtime. Activities become computational nodes, and connections between activities define possible transitions. Process variables form shared workflow state: one activity writes a result, and a later activity or decision reads it.

An exclusive gateway—the diagram equivalent of choosing one route at a junction—evaluates conditions over that state to select a path. Parallel gateways map to branching and synchronization mechanisms intended to coordinate work happening along multiple paths.

An agent-backed activity can be configured with a prompt, a language model and available tools. It receives workflow inputs, may retrieve information or call tools, and returns outputs to the process state. The surrounding model still controls how execution proceeds.

The authors argue that this arrangement gives developers and operators a common view of intended workflow structure and execution progress. That could help locate failures, but the paper does not establish a measured debugging or reliability advantage.

Making a whole process callable

The central composition mechanism is what the authors call an Agentic Participant. Unlike an agent-backed task, which is one activity inside a process, a participant represents a reusable capability that another agent can invoke as a tool.

Behind that interface might sit a single agent with its own instructions and tools. Alternatively, the participant can have an associated BPMN process containing multiple activities, software services, human tasks and further agents. The caller uses the capability interface without needing to understand those internal steps.

This enables “recursive capability composition”: a process can invoke an agent, that agent can invoke another participant, and that participant can execute another process. Here, recursive describes nested capability calls—not evidence that agents learn new abilities or automatically improve themselves.

The paper also describes exposing these capabilities through the Model Context Protocol, or MCP, so external agents could discover and invoke them. MCP supplies the interface; the process model supplies the workflow behind it.

Organizational context, not just instructions

The same models can also provide information about how an organization operates. Beyond activity order, they can describe responsibilities, business objects and relationships between processes.

The authors distinguish process context, such as the current activity and its inputs; domain context, such as organizational terminology and relationships; and runtime context, including current variables, conversation and tool results.

For example, understanding an activity called “Handle Quality Return” may require knowing how customers, deliveries, materials and inspections relate. Semantic models can represent those connections. The proposed benefit depends on that organizational knowledge being complete, relevant and retrievable.

Feasibility is not a correctness guarantee

The evidence is an architectural description and an implementation demonstrating feasibility, not a systematic large-scale evaluation. Three authors are affiliated with Semtation GmbH, the company behind SemTalk; Jan Fillies is affiliated with FU Berlin.

The authors explicitly identify unresolved questions about whether the transformation preserves BPMN behavior across conditional routing, parallel paths, state updates and human interaction. Formal validation remains future work.

A visible workflow also cannot guarantee that an agent's answer is correct or its tool choice appropriate. Callable processes introduce authorization, auditability and execution-control requirements, especially when they can affect business systems or people.

The proposal is therefore a concrete architecture for combining explicit process control with flexible AI behavior—not demonstrated proof that the resulting workflows are safer, faster or more reliable.