Camunda

Overview

Camunda is a process orchestration platform. Organisations draw their business processes as diagrams in BPMN, an international standard notation for workflows, and Camunda runs those diagrams as real, executable processes. It assigns tasks to people, calls systems and APIs, waits for messages and deadlines, applies business rules from decision tables and records every step of every case.

Since 2025 Camunda has also placed AI agents inside these processes. An agent can handle a step that needs judgement, such as reading a document or choosing which checks to run, while the diagram around it sets the limits: what the agent may decide, which tools it may use and when a person must review.

A railway timetable and signalling system is a useful picture. The timetable defines routes, stops and connections. Signals make sure trains move only when it is safe, and every movement is logged. Camunda is the timetable and signalling for business processes. AI agents are a new kind of train that can choose some of their own stops, but only on the tracks and under the signals the process allows.

What it is

Camunda is a set of components around a process engine. The current generation, Camunda 8, is built on an engine called Zeebe.

Where it came from

Jakob Freund and Bernd Rücker founded Camunda in Berlin in 2008, first as a consultancy for business process management. The underlying standards came from the Object Management Group: BPMN 2.0, which made process diagrams executable, and later DMN for decision tables.

Camunda's own open-source engine began in 2013, when it split its development from the Activiti project. This became the Camunda Platform, later called Camunda 7, released under the Apache 2.0 licence and widely embedded in Java applications.

To scale beyond a single relational database, Camunda built Zeebe, a distributed, cloud-native engine. Camunda Platform 8, built on Zeebe, was released in April 2022, alongside a managed SaaS offering. Since October 2024, Camunda 8 self-managed components have been released under the Camunda License, which allows source access and free use for development and testing but requires a commercial licence for production. Camunda has wound down the community edition of Camunda 7, and several community forks now maintain that code line.

In 2025 Camunda introduced agentic orchestration. It added an AI Agent connector that lets a model reason and call tools within a process, and uses BPMN ad-hoc sub-processes as the agent's toolbox: a set of tasks the agent may run in any order it chooses, inside a defined boundary. Camunda 8.8 was released in October 2025.

How it works

Model the process

A process model is a BPMN diagram. Its main building blocks are:

• Service tasks, which call systems through job workers or connectors.
• User tasks, which appear in a person's task list.
• Business rule tasks, which evaluate a DMN table.
• Gateways, which split or join paths based on data.
• Events, such as a timer, an incoming message or an error.
• Sub-processes, which group steps; an ad-hoc sub-process leaves the order open.

Execute instances

Each new case, such as one customer application, becomes a process instance. The engine moves it through the model, keeps its variables and records its history. When it reaches a service task, it creates a job; a worker or connector picks it up, does the work and reports the result. When something fails, the engine retries according to the model, then raises an incident that an operator can fix in Operate.

Make decisions transparent

Rules that people need to read and change, such as eligibility or pricing tiers, sit in DMN tables. Each evaluation is recorded, so auditors can see which rule fired and why.

Place agents inside the model

An AI agent step is part of the diagram. The modeller decides the prompt, the model, the tools the agent may call (other tasks in an ad-hoc sub-process) and what happens with low-confidence results, for example routing to a user task. The process records the agent's inputs, tool calls and outputs like any other step. See human in the loop and AI governance.

Key concepts and terms

A worked example

This is an illustrative scenario. A mobile operator onboards about 4,000 new business customers a month.

1. Start. A message event starts an instance when an online application arrives.
2. Credit. A service task calls the credit bureau through a REST connector.
3. Plan tier. A business rule task evaluates a DMN table using credit score, company size and region. The table has 18 rows that the commercial team can read and change.
4. Documents. An AI agent step reads the uploaded registration documents and extracts company name, registration number and directors. It can call two tools inside an ad-hoc sub-process: a company-registry lookup and a sanctions-list check.
5. Confidence. If the agent's extracted name does not match the registry, or the sanctions tool returns a possible match, a gateway routes the case to a user task for a compliance clerk. In this scenario about 12 percent of cases go to a person.
6. Activate. A final service task activates the lines and sends a welcome email.
7. Deadline. A timer event escalates any case still open after three working days.

Managers use Optimize to see that the clerk review step is the slowest, and auditors can trace every decision, including each tool the agent called.

Strengths and limitations

Strengths:

• Visual models shared by business and IT, based on open standards.
• Strong audit trails for every instance, decision and agent step.
• Built-in handling of people, deadlines, messages and incidents.
• Transparent rules through DMN tables.
• A governed, visible place for AI agents inside regulated processes.

Limitations:

• Good modelling takes skill and discipline.
• Heavier than code-first tools for small or purely technical flows.
• Production use of Camunda 8 self-managed needs a commercial licence.
• Moving from Camunda 7 to Camunda 8 is a real migration project.
• Very open-ended agent work fits poorly into a fixed diagram.

How it compares with alternatives

Practical guidance and best practices

1. Model the process first, and place agents only where judgement adds clear value.
2. Keep business rules in DMN tables so people can read and change them.
3. Give each agent a small, explicit set of tools in an ad-hoc sub-process.
4. Route low-confidence or high-risk agent results to user tasks.
5. Use timer events for deadlines and escalation.
6. Monitor incidents and cycle times per step, and fix the slowest steps first.
7. Version process models, test them on sample cases and deploy through a pipeline.
8. Use compensation events for steps that must be undone on failure; see the saga pattern.

Common mistakes

• Drawing every technical detail into the diagram until nobody can read it.
• Hiding business rules in code instead of DMN tables.
• Putting an AI agent where a simple rule would do.
• Letting agent output flow straight into irreversible steps without a check.
• Ignoring incidents until a backlog builds up.
• Assuming a Camunda 7 model will run unchanged on Camunda 8.

Frequently asked questions

What is Camunda?

Camunda is a process orchestration platform that runs business processes modelled in BPMN and decisions modelled in DMN. It coordinates people, systems and AI agents, and records every step for monitoring and audit.

What is BPMN?

Business Process Model and Notation is a standard from the Object Management Group for drawing processes. Version 2.0 defines both the visual notation and the execution meaning, so the same diagram can be read by people and run by an engine.

What is Zeebe?

Zeebe is the distributed process engine at the heart of Camunda 8. It was built to run large numbers of process instances without depending on a single central relational database.

How do AI agents fit into Camunda?

As steps inside modelled processes. The diagram sets the agent's prompt, model and allowed tools, often as tasks in an ad-hoc sub-process, and defines what happens to uncertain results, such as sending them to a person.

Is Camunda open source?

Camunda 7 was released under the Apache 2.0 licence. Camunda 8 self-managed components are source-available under the Camunda License, free for development and testing, with a commercial licence required for production. A SaaS version is also offered.

When should I choose Camunda over a code-first engine?

When business stakeholders need to see and agree the process, when audit and compliance matter, and when many cases involve people, deadlines and rules as well as systems.

Glossary

• BPMN: a standard notation for executable process diagrams.
• DMN: a standard for decision tables.
• FEEL: the expression language for rules and conditions.
• Instance: one running case of a process.
• Job: a unit of work created by a service task.
• Connector: a ready-made integration used in a model.
• Incident: a failure needing operator action.
• Tasklist: the application where people complete user tasks.
• Operate: the tool for monitoring and fixing running instances.
• Ad-hoc sub-process: a set of tasks that can run in any order.

Summary

Camunda runs business processes drawn in standard BPMN, with decisions in DMN, coordinating people, systems and, since 2025, AI agents inside governed steps. Founded in Berlin in 2008, it moved from the embeddable Camunda 7 engine to the distributed, Zeebe-based Camunda 8. Its strengths are shared visual models, transparent rules and complete audit trails. Model the process first, keep rules readable, give agents small tool sets, route uncertain results to people and watch incidents and cycle times.

Related

It belongs to Workflow automation in my stack, under Agentic AI. Others in the same group: