Camunda 8 (Agentic Orchestration)
Type: low-code · Vendor: Camunda · Language: Java · License: proprietary · Status: active · Status in practice: emerging · First released: 2022
Camunda 8 orchestrates LLM-driven agents inside BPMN processes so an agent decides which tools to call while the Zeebe engine deterministically executes the selected activities, stores state, and routes human tasks.
Description. Camunda 8 is a process-orchestration platform built on the Zeebe engine. Its agentic orchestration runs an LLM agent inside a BPMN ad-hoc sub-process whose activities define the available tools; the model chooses the next tool and the engine executes it, persists variables, and applies retries and incident handling. The same engine runs deterministic flows, decision tables, human tasks, and agentic flows together. Teams use it to embed agents in audited, end-to-end business processes.
Agent loop shape. A user prompt enters an ad-hoc sub-process. The LLM evaluates the prompt, system prompt, and tool definitions and selects which BPMN activity to invoke next; Zeebe executes that activity, stores variables, applies retries and incident handling, and feeds the result back. The loop repeats until the agent produces a final response, with human tasks and events routed by the engine where the model defers to a person.
Primary use cases
- embedding LLM agents in regulated business processes
- orchestrating tool calls as BPMN activities
- mixing deterministic flows, decision tables, and agentic decisions in one engine
- routing agent work to human approvers within a process
Key concepts
- Ad-hoc sub-process → bpmn-dmn-deterministic-shell (docs) — A BPMN embedded sub-process marked ad-hoc whose contained activities have no fixed sequence flow; the AI agent treats those activities as the selectable tool set and the engine activates whichever one the model chooses next.
- AI Agent connector feedback loop → react (docs) — The AI Agent connector runs as a feedback loop between the LLM and Camunda: the model selects a tool, the engine executes the corresponding BPMN activity and stores variables, and the result is fed back to the model until a final response is produced.
- Short-term conversational memory → short-term-memory (docs) — Per-process-instance memory that retains the running conversation so the agent supports multi-turn interactions and follow-up questions inside one process instance.
- MCP tool connection → mcp (docs) — A way to attach the AI Agent connector to tools served by Model Context Protocol servers, extending the agent's tool palette beyond locally modelled BPMN activities.
Patterns this low-code implements —
- ★BPMN/DMN Deterministic Shell Around Agent
BPMN is the deterministic spine: the LLM agent runs inside a BPMN ad-hoc sub-process whose nodes define the available tools/steps. The model only decides which tool to call next; Zeebe deterministica…
- ★★Tool Use
The agent's actions are typed tool calls, not free text: BPMN activities inside the ad-hoc sub-process are exposed as the tool set, the model evaluates the prompt and the available tool definitions,…
- ★★Human-in-the-Loop
Because the agent runs inside a BPMN process, a human task is a native activity the orchestration routes to: agentic decisions, deterministic rules, and human tasks collaborate in the same end-to-end…
- ★★Model Context Protocol
Beyond BPMN activities, the AI Agent connector can call tools exposed by Model Context Protocol servers, so a tool written once for MCP is callable from inside a Camunda process without a bespoke ada…
- ★★Short-Term Thread Memory
The connector carries the conversation slice across loop iterations within the running process instance, so the agent can handle multi-turn interactions and follow-up questions rather than treating e…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
- full-code · framework
Temporal
Temporal is also a durable execution engine used to host agent loops, but it is a code-first workflow runtime rather than a BPMN/DMN low-code modeller.
- low-code · visual builder
n8n
n8n is a low-code workflow platform that likewise embeds LLM agent steps in visual flows, targeting a lighter-weight automation audience than Camunda's process-orchestration spine.
- full-code · framework
Restack
Restack is another durable workflow engine positioned for agent orchestration, sharing the deterministic-engine-around-the-LLM shape without BPMN modelling.
References
Agentic orchestration | Camunda 8 Docs
“Camunda agentic orchestration allows you to orchestrate AI agents within your BPMN-based workflows, enabling human tasks, deterministic rule sets, and AI-driven decisions to collaborate in a robust, end-to-end process.”
“Exposes BPMN activities inside an ad-hoc sub-process as tools the LLM can select.”
“Having both deterministic and dynamic flows run in a single engine enables you to handle the entire spectrum of business processes”
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial