new-agent

name: new-agent description: Generate a Camunda 8 agentic process with AI Agent connector and ad-hoc sub-process. Use when building AI agent workflows with tool calling. argument-hint: "[agent purpose and tools]"

Generate an agentic Camunda 8 BPMN process. Find and read the "Agentic BPMN" section of camunda-dev-guide.md for ad-hoc sub-process XML, connector task types, input targets, and fromAi() usage.

$ARGUMENTS

Use the AI Agent Sub-process pattern (recommended):

1. Create a bpmn:adHocSubProcess with zeebe:taskDefinition type="io.camunda.agenticai:aiagent-job-worker:1"
2. Add zeebe:adHoc outputCollection="toolCallResults" outputElement="={id: toolCall._meta.id, name: toolCall._meta.name, content: toolCallResult}" — required for tool result collection
3. Add zeebe:taskHeaders with resultVariable = agent
4. Add tools as activities inside (no incoming flows, no start/end events)
5. Use fromAi() in tool input mappings for AI-generated parameters
6. Each tool must write output to toolCallResult variable (e.g. resultExpression → ={toolCallResult: response.body})
7. Add <bpmn:documentation> to each tool describing what it does
8. Configure systemPrompt, userPrompt, modelProvider, and authentication.apiKey ({{secrets.KEY_NAME}})

Tools can also be sub-processes for multi-step async patterns:

• Put a service task (action) followed by a catch event (wait) inside a sub-process
• The sub-process is the tool entry point; inner steps are hidden from the agent
• Use this for send+wait, request+callback, or any tool needing async interaction
• See the "Sub-flow Tool Pattern" section in camunda-dev-guide.md

For non-REST connectors as tools, look up their task type and input targets via the Camunda Docs MCP tool.

Save to resources/. Deploy with c8ctl.