Autarq is a contributor to the upcoming F13 release and is developing an agentic component that lets the platform handle multi-step processes for public agencies and municipalities.
TL;DR: Autarq is contributing to the further development of F13, the AI assistant of the State of Baden-Württemberg. At the center of our contribution is an orchestrator that enables agentic work on the platform: F13 will then be able to independently handle more complex, multi-step, and varied tasks, instead of only answering individual requests as before. Human approval points and traceable decision rules are built firmly into the process. The agentic orchestrator is scheduled to launch in the second half of 2026.
F13 has been open source since July 2025. What began as an application of the Baden-Württemberg state administration has become a collaborative project that public administrations, government departments, and companies can contribute to. The state has explicitly invited this participation.
At Autarq, we’re contributing an agentic component to F13. It breaks a case down into individual steps, logs each one, and requests approval at defined points. The professional decision stays with the human.
What is F13, and what does the move to open source mean?
F13 supports state administration employees with writing tasks, offers a chatbot, summarizes documents, and searches administrative documents. A prototype was available to all state administration employees starting in May 2023, followed by the official rollout as a full product in 2024. The software was developed by the state’s innovation lab (InnoLab_bw) together with PD – Berater der öffentlichen Hand GmbH. It runs in the data center of the state’s own IT provider, IT Baden-Württemberg (BITBW).
In July 2025, the state released F13 as open-source software, made available through openCoDE, the platform of the Center for Digital Sovereignty (ZenDiS). Ministerialdirektor Stefan Krebs, the state government’s Commissioner for Information Technology, put the scale into perspective at the announcement: the federal government, sixteen states, and, in Baden-Württemberg alone, 1,101 municipalities all need to prepare for the use of AI.
Technically, F13 consists of individual microservices for chat, research, summarization, and transcription, held together by a core. F13 can be operated in your own data center and is not tied to a specific AI model. New features are added as new microservices instead of rebuilding existing ones. This architecture is the precondition for our contribution to F13.
What exactly is Autarq working on at F13?
In its current feature set, F13 answers questions and processes documents. What’s missing is the ability to handle multi-step workflows, in which the system selects tools, evaluates intermediate results, and carries a case through to a submission-ready outcome. On the public F13 roadmap, this area is planned for 2026 as foundational work on request routing and agents.
This is where our component comes in. At its center is an orchestrator that breaks a task down and distributes it to specialized sub-agents: one researches, one checks against the rule sets, and one drafts. It brings the results together and decides whether a step needs to be repeated. Compared to a single large model call, this keeps every individual step visible and separately reviewable.
For us, human-in-the-loop means AI and people work within the same process, not alongside it. The component handles the groundwork — researching, evaluating, preparing — and presents the result to caseworkers along with its reasoning and sources. At defined points, the system pauses and waits for their approval. Only once a person has reviewed and confirmed that a result is correct and complete does the case move on to the next steps. In public administration, responsibility for a decision is always assigned to a specific person — which is why this approval is built firmly into the process, not optional.
Decision rules live in a policy engine built on Open Policy Agent (OPA) — as standalone, versionable policies, rather than as prompt text or a special case buried in the code. Anyone who wants to know why an agent is allowed to do something can read the policy and review it from a domain perspective, without reading the code. In addition, a log records every agent action: which tool was called, which source was used, and which policy applied. This makes it possible to reconstruct a case after the fact.
Specialist systems and document repositories are connected via the Model Context Protocol (MCP), an open standard. An MCP server built once can be reused by other public administrations and keeps working even if the model changes. A document manager stores documents and case data centrally, so multiple agents can access them.
The orchestrator itself is also built on open standards: for planning agent workflows and controlling sub-agents, we rely on established open-source frameworks. For integration into the F13 interface, we additionally use the AG-UI protocol. All building blocks are designed as interchangeable components in a Docker stack, following the same pattern F13 already sets with its microservices. The component runs in the same environment as F13 — that is, in the data centers of the respective organization.
How does collaboration within the F13 community work?
The F13 community has a fixed process for larger contributions. Ideas are discussed in a biweekly contribution office hour with the lead developer. Technical requirements and architecture decisions then go through two boards that meet every three weeks. The requirements board is made up of the technical product owners of all active projects, and the architecture board of the lead developers. The binding basis for any larger contribution is an EPIC on openCoDE, with business goals, technical guardrails, and clear ownership.
On top of that, there’s a license check that runs automatically in the F13 project’s CI/CD pipeline. If licenses change in a merge request, they’re reviewed legally before the contribution is accepted. Our code is developed open source and REUSE-license-compliant, and goes through the same review as everyone else’s.
Other contributors to F13 include the KIRKE project of the Baden-Württemberg police, the Ministry of Education, Youth and Sport with SCHULE@BW, the state of Saarland, as well as InnoLab_bw and PD.
What’s next?
We’re working iteratively, from concept through prototype to production. That takes longer than building something in isolation, but the result is interoperable and reusable by any public administration. We’ll keep you updated on further developments on our blog and on LinkedIn.
Frequently asked questions
What is an agentic component?
An agent works through a task in multiple steps. It selects tools, retrieves data, evaluates intermediate results, and brings them together. Our component brings this way of working to F13, with defined points in the process that require human approval.
Does the professional decision stay with caseworkers?
Yes. Human-in-the-loop is part of our architecture. The component prepares a case and makes the reasoning and sources transparent. Caseworkers make the professional decision.
How are the agents' decisions made traceable?
Through two mechanisms. Decision rules exist as standalone policies in Open Policy Agent, so they can be reviewed from a domain perspective without reading the code. In addition, every agent action is logged: the tool called, the source used, and the policy that applied.
What is MCP used for?
The Model Context Protocol is an open standard for accessing tools and data sources. An MCP server built once can be reused by other public administrations and keeps working even after a model change.
Will the code be published?
Yes. The code is developed open source and REUSE-license-compliant within the F13 project, making it available for reuse by the federal government, states, and municipalities.
Where does the component run?
In data centers in Germany or the EU. The component is built as a Docker stack of interchangeable building blocks and is model-agnostic.
Do you have a process with many manual steps?
Do you work at a public agency or municipality and have a process in mind that today requires many manual steps in between? Get in touch, and let’s talk about what an agentic component could do for your project!
Autarq is the brand of MWAY DIGITAL GmbH from Stuttgart and builds the sovereign managed cloud platform for Europe. Autarq Desk is our sovereign agentic workspace based on openDesk CE, augmented with Autarq/Chat, Nubus IAM, and Autarq Network, optionally extensible with add-ons such as Relution for MDM/UEM. All components run in German data centers, with AGPLv3-compliant stack composition and full tenant isolation at the Kubernetes level.
Trademark notice: F13 is an application of the State of Baden-Württemberg. openCoDE is a platform of the Center for Digital Sovereignty (ZenDiS). Open Policy Agent is a project of the Cloud Native Computing Foundation. All other product and project names mentioned are trademarks or designations of their respective owners and are used here solely for identification purposes.
