Project: Agentic Orchestration Studio · Status: Draft · Date: April 2026
Agentic Orchestration Studio
A shared platform for mapping, automating, orchestrating, and measuring business processes in Aarhus Kommune.
Introduction
Aarhus Kommune is increasingly adopting AI agents, RPA bots, and workflow automations across departments. But these efforts are fragmented — each initiative lives in isolation, with no shared overview, no common tooling, and no way to measure the collective business value.
This project explores whether we can build a shared orchestration platform that gives the organization:
- A visual map of automated processes
- Tools to orchestrate agents, RPA, and workflows together
- Human-in-the-loop routing when decisions need a person
- Dashboards showing where processes are and what value they deliver
- Visibility into dependencies on backend systems — useful for procurement and architecture planning
The inspiration comes from TwoDay's commercial Agentic Orchestration Studio (AOS), presented at OffDig 2025. The question is: can we build this ourselves, using open-source components, tailored to our context, security requirements, and scale?
And beyond replicating what exists — can we go further? Can we have an AI assistant built into the platform that helps users design, refine, and optimize their process flows?
The Problem
Today, automation in the municipality looks like this:
- RPA bots run in one system, managed by one team
- AI agents are developed in another context, with different tooling
- Workflow automations (form processing, case routing) live in yet another silo
- No shared overview — nobody can see all automations in one place
- No measured business value — we can't tell leadership what the return on investment is
- No visibility into human-in-the-loop steps — where do processes need human intervention, and how long does that take?
- No dependency map — when a backend system changes, which automations break?
- Scaling is hard — each new automation is a standalone effort with its own infrastructure
This makes it difficult to prioritize, to scale, and to make a strategic case for further investment.
What TwoDay's AOS Does
TwoDay's Agentic Orchestration Studio is organized around four capabilities:
Map
Visual process mapping — documenting what a process looks like, who's involved, what systems it touches, and where automation is possible.
Automate
Building the automations themselves — connecting AI agents, RPA bots, and integrations into executable workflows.
Orchestrate
Running and managing automations in production — routing tasks between bots and humans, handling exceptions, managing dependencies, and ensuring processes complete end-to-end.
Analyze
Measuring the effect — dashboards showing process throughput, error rates, time savings, and documented business value (ROI).
Key strengths of the AOS approach:
- Everything in one place — mapping, execution, monitoring
- Built-in AI that suggests process optimizations
- Human-in-the-loop as a first-class concept
- Dependency visibility to backend systems
- Business value measurement built in
The limitation: It's a commercial platform. Licensing costs scale with users, creating a potential barrier to broad adoption across 30,000+ municipal employees. And it's a dependency on a single vendor.
The Open-Source Opportunity
Full analysis
See the Open Source Landscape for detailed tool-by-tool evaluation.
Key finding
No single open-source tool covers all four capabilities (Map, Automate, Orchestrate, Analyze). But strong tools exist for each layer, and they can be composed into a platform.
The strongest candidates
| Capability | Best open-source option | Notes |
|---|---|---|
| Process mapping (BPMN) | Flowable + BPMN.io | Industry-standard BPMN 2.0. Flowable also has case management (CMMN) and decision tables (DMN). |
| Agent orchestration | LangGraph or Microsoft Agent Framework | Both MIT-licensed. LangGraph has the largest community. Microsoft's framework just hit 1.0. |
| Workflow automation & integration | n8n or Activepieces | 400–600+ connectors. Visual builders. Self-hosted. |
| RPA | Robot Framework | Mature, Apache 2.0, huge ecosystem. Keyword-driven (accessible). |
| Monitoring & analytics | OpenTelemetry + Grafana | Industry standard. Flexible dashboards. |
Flowable stands out
Flowable is the closest thing to a single platform because it uniquely combines:
- BPMN 2.0 process engine (mapping + execution)
- CMMN case management (adaptive processes)
- DMN decision tables
- Human task management (first-class)
- A new agent engine (2025–2026) for AI agent orchestration alongside BPMN
The open-source edition (Apache 2.0) covers the engines. The visual design studio and AI features are commercially licensed — but the core is solid and self-hostable.
Agent-Assisted Process Design
A differentiator beyond what TwoDay offers: an AI assistant embedded in the platform that helps users create and improve process flows.
What it could do
| Capability | Example |
|---|---|
| Propose flows | "I need to handle citizen parking complaints" → agent generates a BPMN draft with steps, decision points, and human tasks |
| Suggest automations | Agent analyzes a manual process and highlights which steps could be automated, with which tools |
| Identify integrations | "This step needs data from SBSYS" → agent suggests the n8n connector and maps the fields |
| Refine existing flows | Agent reviews a running process and suggests optimizations based on actual execution data (bottlenecks, error rates) |
| Research patterns | "How do other municipalities handle building permits?" → agent searches templates and best practices |
How it would work
The assistant would be a LangGraph agent with access to:
- The BPMN modeler — it can read and generate BPMN 2.0 XML, proposing flows visually
- The integration catalog — it knows which n8n connectors are available and what they do
- Existing process templates — it learns from flows already built on the platform
- Execution analytics — it can read Grafana metrics to identify bottlenecks and suggest improvements
The user interacts via a chat panel alongside the process modeler. The agent proposes, the user approves and refines — keeping the human in control while removing the blank-canvas problem.
Why this matters
Most process mapping tools assume the user already knows what the process should look like. But process owners often don't — they know the outcome they want, not the optimal flow. An AI assistant bridges that gap and dramatically lowers the barrier to entry.
Recommended Architecture
Full details
See Architecture Options for detailed stack descriptions.
Option A: Full composed stack
For maximum flexibility and full coverage of all four capabilities:
| Layer | Tool | Role |
|---|---|---|
| Process mapping & execution | Flowable | BPMN/CMMN/DMN engines, human task management |
| Visual process modeler | BPMN.io | Web-based BPMN editor, embeddable |
| Agent orchestration | LangGraph | AI agent workflows with human-in-the-loop |
| Integration & automation | n8n | 400+ connectors, visual workflow builder |
| RPA | Robot Framework | Desktop/browser/API automation |
| Monitoring & dashboards | Grafana + OpenTelemetry | Process metrics, SLAs, ROI tracking |
Option B: Minimal viable stack
For a faster start with fewer moving parts:
| Layer | Tool |
|---|---|
| Process mapping, execution & human tasks | Flowable |
| Integration & automation | n8n |
| Monitoring | Grafana |
Option B gets us Map + Orchestrate + basic Analyze with three components. Agent orchestration and RPA can be added later as the platform matures.
Why Build vs. Buy
| Factor | Build (open-source) | Buy (TwoDay AOS) |
|---|---|---|
| Licensing cost | Free (open-source core) | Per-user licensing — expensive at scale |
| Customization | Full control, tailored to our context | Limited to vendor's roadmap |
| Security | Self-hosted, full data sovereignty | Depends on deployment model |
| Vendor dependency | None (open standards, replaceable components) | Single vendor lock-in |
| Time to value | Longer — requires development and integration | Shorter — turnkey platform |
| Maintenance | Our responsibility | Vendor-managed |
| Community | Open-source communities, OS2 collaboration | Vendor support |
| Scalability | Unlimited users, no license barriers | License cost scales with users |
Our position: We have the technical capability to build this. The open-source components are mature. And building on open standards means we can share the platform with other municipalities through OS2 — multiplying the value.
The risk is development effort and maintenance. A phased approach (start minimal, grow based on proven value) mitigates this.
Open Questions
Full details
See Considerations for all open questions and discussion points.
Key questions to resolve:
- [ ] Target audience — who uses this? IT only, or also process owners and business units?
- [ ] Roles and permissions — who can map, who can orchestrate, who can just monitor?
- [ ] OS2Forms 2.0 — is this the future form system, and should our platform integrate or replace parts of it?
- [ ] Existing systems — what integrations are needed on day one? (KMD, SBSYS, etc.)
- [ ] Security model — classification of data flowing through the platform
- [ ] Funding — what's the investment case? Can we quantify the cost of fragmentation?
- [ ] OS2 collaboration — interest from other municipalities in co-developing?
Next Steps
- Validate the concept — present this research to stakeholders, test appetite
- Map current automations — inventory what's already running across departments
- Proof of concept — stand up the minimal stack (Flowable + n8n + Grafana) with one real process
- Create interactive mocks — UI prototypes for the "dashboard" and "process map" views to sell the vision
- Engage OS2 — explore co-development with other municipalities
- Business case — quantify the cost of fragmentation vs. the cost of building
Sources
- TwoDay AOS: twoday.com/services/applied-ai/aos
- OffDig presentation: "Fra AI-eksperimenter til dokumenteret forretningsværdi" (March 2025)
- See Open Source Landscape for all tool-specific sources