Skip to content
Last updated: 23. apr. 2026, 06.10

Project: Agentic Orchestration Studio

Architecture Options

Two proposed architectures for an open-source orchestration platform, from comprehensive to minimal.


Option A: Full Composed Stack

Maximum flexibility. Covers all four capabilities (Map, Automate, Orchestrate, Analyze) with best-of-breed tools at each layer.

Components

LayerToolRole
Process engineFlowableBPMN 2.0 process execution, CMMN case management, DMN decision tables, human task management
Process modelerBPMN.ioWeb-based visual BPMN editor, embeddable in our UI
Agent orchestrationLangGraphAI agent workflows — classification, summarization, decision support, with human-in-the-loop
Integration & automationn8n400+ connectors to external systems (forms, case management, email, APIs)
RPARobot FrameworkDesktop/browser/API automation for legacy systems
Operations monitoringGrafana + OpenTelemetryInfrastructure health, alerting, detailed analytics (IT/admin only)
Data storePostgreSQLShared database for process state, audit logs, analytics
IdentityKeycloakSingle sign-on, role-based access, municipal AD integration

How it fits together

┌─────────────────────────────────────────────────────────┐
│                    Web Frontend                          │
│        (Process maps, dashboards, task inbox)            │
├────────────┬────────────┬────────────┬──────────────────┤
│  BPMN.io   │  Task UI   │  Dashboard │   Admin          │
│  (modeler) │  (inbox)   │  (custom)  │   (config)       │
├────────────┴────────────┴────────────┴──────────────────┤
│                   API Gateway                            │
├────────────┬────────────┬────────────┬──────────────────┤
│  Flowable  │  LangGraph │    n8n     │ Robot Framework  │
│  (BPMN/    │  (agents)  │  (integr.) │ (RPA)            │
│   CMMN/    │            │            │                  │
│   DMN)     │            │            │                  │
├────────────┴────────────┴────────────┴──────────────────┤
│              Keycloak (identity/SSO)                     │
├─────────────────────────────────────────────────────────┤
│    PostgreSQL    │    OpenTelemetry    │    Grafana      │
│    (state/audit) │    (telemetry)      │    (dashboards) │
└─────────────────────────────────────────────────────────┘

            ┌─────────────┼─────────────┐
            │             │             │
        ┌───┴───┐   ┌────┴───┐   ┌────┴───┐
        │ KMD   │   │ SBSYS  │   │OS2Forms│
        │       │   │        │   │        │
        └───────┘   └────────┘   └────────┘
         (external systems / backends)

Interaction patterns

  1. Process designer uses BPMN.io to model a process visually
  2. Flowable executes the BPMN process, routing between automated steps and human tasks
  3. At an "agent step", Flowable delegates to LangGraph which runs the AI agent(s)
  4. At an "integration step", Flowable triggers n8n workflows to talk to external systems
  5. At an "RPA step", Flowable triggers a Robot Framework job for legacy system automation
  6. Human-in-the-loop tasks appear in a task inbox UI, powered by Flowable's task service
  7. The platform dashboard (custom, built into the web frontend) shows process stats, task counts, hours saved, and ROI to all users
  8. OpenTelemetry collects infrastructure metrics from all components; Grafana provides detailed operational dashboards, alerting, and ad-hoc analytics for IT/admins
  9. Keycloak handles authentication and role-based access across all components

Strengths

  • Full coverage of Map, Automate, Orchestrate, Analyze
  • Each component is replaceable (no vendor lock-in)
  • Scales independently per layer
  • Standards-based (BPMN 2.0, OpenTelemetry, OIDC)

Risks

  • Integration complexity — 6+ components to connect and maintain
  • Requires strong DevOps capability
  • More moving parts = more potential failure points
  • Longer time to first value

Option B: Minimal Viable Stack

Fastest path to a working platform. Three components covering the core needs.

See the mock

The Unified Platform mock ↗ demonstrates how these three components can feel like a single seamless platform.

Components

LayerToolRole
Process engine + human tasksFlowableBPMN execution, case management, task inbox
Integration & automationn8nExternal system connectors, simple automations
Platform dashboardCustom (built-in)Process stats, task overview, hours saved — for all users
Operations monitoringGrafana + OpenTelemetryInfrastructure health, alerting, ad-hoc analytics — for IT/admins

How it fits together

┌─────────────────────────────────────┐
│            Web Frontend              │
│    (Process maps, task inbox,        │
│     dashboards)                      │
├──────────────┬──────────────────────┤
│   Flowable   │        n8n           │
│  (BPMN +     │   (integrations      │
│   tasks +    │    + automations)    │
│   dashboard) │                      │
├──────────────┴──────────────────────┤
│     PostgreSQL    │  OpenTelemetry  │
├───────────────────┤  + Grafana     │
│   (state/audit)   │  (ops only)    │
└───────────────────┴─────────────────┘

        ┌─────────┼─────────┐
        │         │         │
    ┌───┴──┐  ┌──┴───┐  ┌──┴────┐
    │ KMD  │  │SBSYS │  │OS2Forms│
    └──────┘  └──────┘  └───────┘

What you get

  • Map: BPMN process modeling via Flowable's built-in modeler (or BPMN.io added later)
  • Automate: n8n handles integrations and simple automations
  • Orchestrate: Flowable manages process execution and human-in-the-loop
  • Analyze: Platform dashboard shows process stats and ROI to all users; Grafana provides operational monitoring for IT

What you don't get (yet)

  • AI agent orchestration (add LangGraph when ready)
  • RPA for legacy desktop apps (add Robot Framework when needed)
  • Advanced analytics / ROI measurement (evolve the platform dashboard)

Strengths

  • Three components — manageable to operate
  • Fast to stand up (weeks, not months)
  • Proves the concept before investing in the full stack
  • Each component has a large, active community

Growth path

Minimal stack (Phase 1)
  └─ + LangGraph (Phase 2: AI agents)
       └─ + Robot Framework (Phase 3: RPA)
            └─ + Keycloak (Phase 4: SSO/roles)
                 └─ + BPMN.io (Phase 5: visual modeler)

Two Dashboard Layers

An important distinction: the platform has two kinds of dashboards serving different audiences.

Platform dashboard (built into the app)

This is what all users see — process owners, caseworkers, managers. It shows:

  • Active processes, completion rates, task counts
  • Hours saved, estimated ROI
  • Process bottlenecks (where are things stuck?)
  • Personal task overview

This is a custom-built part of the web frontend, pulling data from Flowable's process engine and PostgreSQL. It does not require Grafana.

Operations dashboard (Grafana — IT/admins only)

This is where IT monitors the platform itself. It shows:

  • Infrastructure health — is Flowable responding? n8n queue depth? Database connections?
  • Error rates and alerting — "n8n automation X has failed 3 times in the last hour"
  • Detailed analytics — ad-hoc queries like "average completion time for building permits by month over the last year"
  • SLA monitoring — "3 tasks have breached their 48-hour SLA"
  • Resource utilization — CPU, memory, disk across all components

Grafana connects to OpenTelemetry (for infrastructure metrics) and PostgreSQL (for process data). It's accessible via a separate URL or linked from an admin section in the platform.

Platform dashboardGrafana (ops)
AudienceEveryoneIT / admins
PurposeBusiness value, task overviewInfrastructure health, alerting
Data sourceFlowable + PostgreSQLOpenTelemetry + PostgreSQL
Built byUs (part of the app)Configuration (Grafana dashboards)
AccessMain navigationAdmin section / separate URL

Deployment

Both options should be deployed as containers (Docker/Kubernetes) for consistency and scalability. All components support containerized deployment.

Infrastructure requirements (minimal stack)

ComponentCPUMemoryStorage
Flowable2 cores4 GB10 GB
n8n1 core2 GB5 GB
Grafana1 core1 GB5 GB
PostgreSQL2 cores4 GB50 GB
Total6 cores11 GB70 GB

This fits comfortably on a single server or a small Kubernetes cluster.


Recommendation

Start with Option B. Get a working platform with one real process running end-to-end. Use it to demonstrate value and build organizational buy-in. Then grow toward Option A based on actual needs, not assumptions.