AI agentic development
We build AI agents into governed processes. An agent runs as a step of a Camunda BPMN process, through the Camunda AI Agent connector. The process decides when it acts, which tools it may call and when a person approves. We also prepare the data the agents work with.
The task.
An AI agent that acts on its own is hard to audit. You need to know what it saw, what it decided and who approved it. A BPMN process can hold those answers.
The agent also reads your data. If that data is wrong or incomplete, so is the agent’s answer. If it is not cleared for this use, the agent must not see it. The data work comes before the agent.
What we do.
- 01
The agent as a process step
The Camunda AI Agent connector runs the agent inside the BPMN model. The process starts it, gives it its context and takes its result.
- 02
Tools the model defines
The agent’s tools are BPMN activities in an ad-hoc sub-process. The agent chooses among them; the model decides which exist.
- 03
Rules and approvals
Decisions that must not vary stay in DMN. Where a person must approve, a user task waits in Tasklist.
- 04
The audit trail
Each process instance keeps its history: the inputs, the agent’s result, the approvals and the path taken. Operate shows that history for each instance.
- 05
The data behind the agents
We map which data each agent may read, clean it and set its access rules. This part is done with Vyara Tonkova, PhD, our Senior Data Scientist, and our data engineers.
- 06
Tests before go-live
We test the agent against cases with known answers before it goes live. After go-live, its results are compared with the outcomes the process records.
Typical engagements.
First agent in a process
One agent added to one existing Camunda process, with its tools, rules, approvals and tests.
Data preparation
The data an agent will read, mapped, cleaned and put under access rules before the agent is built.
Agent design review
A written review of a planned agent: what it decides, what the process decides and what the audit trail records.
The dates.
From EUR-Lex and the Commission's pages, checked on 18 October 2026.
Agents run on Camunda 8; the current release is 8.10, checked on 17 October 2026.
Under the EU AI Act, the transparency duties of Art. 50 apply from 2 August 2026.
The rules for high-risk AI systems in Annex III apply from 2 December 2027, as moved by the AI Omnibus.
We checked the AI Act dates on 18 October 2026.
Is this for you?
A good fit
- You have Camunda 8, or plan to, and want an agent inside a process.
- You must show what an agent saw, what it decided and who approved it.
- Your data owners can say which data an agent may use.
Not a fit
- You want an agent that acts outside any process, rule or audit trail.
- You need legal advice on the AI Act: that is your legal team’s work.
- You need people placed, with no technical lead on your side.
Questions.
How do AI agents fit into a Camunda process?
An AI agent runs as one step of a BPMN process, through the Camunda AI Agent connector. The process decides when the agent acts, which tools it may call and when a person approves.
Who stays in charge, the agent or the process?
The process stays in charge: the BPMN model defines the steps, the tools and the approvals. The agent chooses only within what the model allows, and the instance history records each step.
What data does the agent see?
An agent sees only the data the process passes to it and what its tools return. We map that data with your data owners before the agent is built.
Do you also handle AI compliance?
Yes, our AI compliance work covers the EU AI Act, GDPR, ISO/IEC 42001 and DORA. It is technical and governance work beside your legal team, not legal advice.
Which version are you on?
Tell us the platform, the version and what has to change. We reply with questions or a proposed next step.
Talk to an architect