It’s 9 am at the insurance company and your application architect watches the backlog creep up again. Her team wants an AI task agent to triage emails, pull customer context, and propose next steps. The issue is not whether an LLM can help, but how to model the agent so it runs reliably and scales, while keeping governance and an audit trail intact.
Camunda provides multiple methods so you can decide how you want to implement AI agents within your BPMN process model and how much control you want to include. In fact, you determine what and how much tool-calling the agent handles internally versus what you model explicitly. That choice drives clarity for auditors, cost controls for FinOps, and guardrails for risk teams.
Why does this matter?
Incorporating AI into your processes allows you to blend both deterministic and dynamic orchestration, automating work that could not be automated in the past. In addition, AI shortens cycles and lifts quality while maintaining control and governance over the process.
Camunda gives you two different connectors that support three practical patterns supporting AI in your processes. You can move fast with a lean model, or you can expose every tool call as a first-class step for compliance and cost tracking. Both approaches work. Your KPIs and risk posture decide which one you start with.
The two connectors
As mentioned, Camunda provides two AI agent connectors that support the patterns that you need to meet your AI challenges and requirements.
- The AI Agent Task connector executes a single AI-powered action with tool calling capabilities. You can think of this like a smart LLM call that you can wrap with BPMN for control, guardrails and oversight.
- The AI Agent Sub-process connector runs a multi-step AI reasoning loop with dynamic tool selection. It uses the ad-hoc sub-process implemented by a Zeebe job worker so that the agent can pick and loop across tools internally.
The supported patterns
These two connectors open the possibilities to use three different patterns to implement AI in your BPMN model. You can implement the pattern you need to support your requirements using one or both of these connectors in a single process.
The three patterns are:
- AI Agent Task connector alone
This pattern is a single AI Agent Task connector task that calls an LLM with no extra tools. - AI Agent Task connector with loop and ad-hoc sub-process
This pattern also uses the AI Agent Task connector but adds an ad-hoc sub-process as a multi-instance toolbox that contains your tools. The feedback loop allows the connector to be called repeatedly based on the tool call results until the final goal is achieved. - AI Agent Sub-process alone
This pattern uses the AI Agent Sub-process connector. The brains for this connector are placed at the sub-process connector task with at least one activity inside the sub-process. The job worker implements the agent’s reasoning and tool loop internally and chooses which tool to call next to achieve the goal.
In this blog series, we will describe each of these patterns, when to use them, strengths and trade-offs and provide an example of each pattern. The series will also include a comparison chart with best use cases, strengths, and weaknesses.
Get started now on next blog in the series, where we discuss the first pattern.
Start the discussion at forum.camunda.io