Camunda

Agentic Orchestration, Announcements, Community, Product, Releases

Camunda 8.10 Webinar Recap

A recap of Camunda 8.10's webinar: build re-engineered processes in days, prove AI agents are trustworthy, and scale reuse across every team.

By Michael Nketsiah

Key Takeaways

  • Camunda 8.10 adds four ways to build a process: ProcessOS (Early Access, outcome-first), Agentic (coding agent + c8ctl), IDE (Java/C# SDKs, CPT), and Hub (visual, browser-based).
  • The new MCP start event template exposes any deployed Camunda process as a tool any MCP client or AI agent can call.
  • Camunda Process Test now grades AI agent behavior on meaning and reasoning, not exact text match, using similarity checks and judge-model scoring.
  • Operate shows live AI agent state and decision trails; Optimize adds an agentic control plane tracking token usage, reliability and duration over time.
  • Camunda Hub merges Web Modeler and Console into one app, organized as Organization → Workspace → Project, with a shared catalog of reusable element templates.
  • 8.10 introduces Physical Tenant multi-tenancy: one cluster, isolated data stores, identity and backup per tenant, between logical tenancy and full multi-cluster separation.
  • ProcessOS is in Early Access for Enterprise customers and certified partners (SaaS and Self-Managed); general availability and full Enterprise-package inclusion are planned for later.

Every process is legacy now

Camunda's October 8.10 webinar opened with a strong claim: every process running in your business today is legacy. Not because it's old. Because someone designed it before AI agents existed, and that design can't hold them now.

The fix runs through three pillars: Speed, Trust, Scale. The team walked through all three using one running example: Camundanzia Financial, a fictional bank, and a personal loan for a customer named Elena Marku.

Elena is a politically exposed person, so her loan needs a real KYC check, not a rubber stamp. Lending takes her application, Financial Crime runs the check with an AI agent and a human reviewer, Lending assesses credit, Payments releases the funds. That one loan, crossing three teams, is what each section below builds, proves, and scales.

Speakers:

• Bastian Körber (SVP Product)

• Aleksander Dytko (Senior Product Manager)

• Felix Mueller (Senior Principal Product Manager)

• Tally Eting (Principal Product Manager)

• Michael Nketsiah (Senior Solutions Marketing Manager), Host

Watch the full recording

Speed: from months to days

Four places consume the time in any re-engineering project: understanding how a process runs today, building it right, proving it works, and connecting it to whoever needs it. Camunda 8.10 reduces each one.

Four places re-engineering loses time, and the 8.10 tool that closes each gap.
Four places re-engineering loses time, and the 8.10 tool that closes each gap.

Four ways in, one way to test:

• ProcessOS (Early Access): hand it files, systems, or web pages describing today's process. It scopes, discovers the as-is, transforms it into an agentic to-be, and implements BPMN, DMN, forms, and workers, with your experts signing off at every stage.

• Agentic: your coding agent builds and deploys using Camunda skills, MCP, and the c8ctl CLI. In Camunda's own demo, the agent took a one-page brief to a deployed loan process in about two minutes.

• IDE: developers code against the Java and C# SDKs, model in Desktop Modeler with reusable credentials, and test with Camunda Process Test (CPT).

• Hub: model, configure and test visually in the browser.

Whichever path a team picks, the tests stay the same. Test Studio records a manual run as a test and saves it as a reusable test scenario, so that same test runs in CI through Camunda Process Test without rebuilding it. Once a process is deployed, the new MCP start event template turns it into a tool any MCP client or AI agent can call, with the same permissions and audit trail as the REST API.

More on this: Camunda Hub governance blog, Camunda Design System blog

Trust: the harder question

An agent can build a process in two minutes. Would you let it make decisions about your customers? That's the question Camunda built this section to answer: not whether agents are capable, but whether you can trust them.

Four questions, four moments, one answer each: Camunda Process Test, Operate and Optimize.
Four questions, four moments, one answer each: Camunda Process Test, Operate and Optimize.

Four people ask four questions, at four different moments, and 8.10 answers all of them in the tools they already use:

• Before go-live, developers ask how it will behave. Camunda Process Test now asserts on meaning, not exact wording: a similarity check on what an answer says, a judge model scoring the reasoning against a threshold you set, and a normal process assertion confirming the case landed on the right task.

• While it runs, operators ask what it's doing. Operate shows the agent's state, its tools in flight, and usage against your limits, for any agent that reports through the Agent Instance API, including Camunda's own agents and frameworks like LangGraph.

• When complete, compliance asks why it did that. Every agent run generates a decision trail: prompt, reasoning, tools called, results, and more. What used to take days of digging through logs now takes minutes.

• Months later, process owners ask if it's still performing. Optimize's new agentic control plane tracks token usage, reliability and duration per process, compares versions, and flags outliers with per-step heatmaps.

In the demo, a new model version brought the KYC agent's failure rate down but raised the average token cost from about 26,000 to about 35,000 tokens per run. The new version also changed tools and prompts, so the model update alone doesn't explain the result. The point stands either way: you decide that trade-off on numbers, not a feeling.

Scale: one hub, governed reuse

One team shipping fast is a start. Twenty teams each rebuilding the same KYC check in isolation is how automation grinds to a halt.

Camunda Hub merges Web Modeler and Console into one application: one login, one place to model, deploy, and manage clusters. It's structured as Organization, then Workspace, then Project. Lending, Financial Crime and Payments each get their own workspace, with their own members and environments. Hub manages clusters, but it never sits in the runtime path, so if Hub goes down, processes keep running.

The Hub catalog is where reuse actually happens: element templates, approved once by the CoE, that apply to connectors, forms, DMN-calling tasks and call activities, and get reused by every team. Lending's developers search the catalog, find “Verify Identity (KYC)” and the sanctions screening that Financial Crime already built, and drop them into the model instead of rebuilding them.

Logical, physical or multi-cluster: same platform, different levels of isolation for different banking needs.
Logical, physical or multi-cluster: same platform, different levels of isolation for different banking needs.

For teams that need to share infrastructure without losing isolation, 8.10 adds a Physical Tenant model, sitting between lightweight logical tenancy and fully separate clusters: one shared cluster, isolated data stores, and backup and identity per tenant. The new Business Value Dashboard lets process owners set cycle time and automation targets per process and see who's hitting them.

More on this: multi-region deployments blog, cross-region disaster recovery blog

Watch the full recording to see Camundanzia’s loan process built, tested and scaled end to end.

Top questions from the webinar

You don't have to choose between a locked-down production network and a shared place to design and govern processes. Here's how Camunda 8.10 handles that, plus the other questions that came up most.

Is Camunda 8.10 lighter?
Since 8.8, the Orchestration Cluster combines broker, gateway, Operate, Tasklist and Identity into one application, so there's less to deploy and operate. 8.10 also simplifies RocksDB memory sizing to a fraction of system memory rather than a fixed amount per partition.

How does Hub work with isolated or locked-down environments?
Hub runs in its own environment and registers your runtime clusters; it doesn't sit inside dev or production. For a fully locked-down production network, Camunda supports a pre-production Hub for modeling and collaboration, and a separate production Hub inside the production network for operations and reporting. Registering a cluster doesn't open data flow between environments; only a limited, allow-listed set of APIs feeds the dashboards.

Can I trigger a Camunda process from Azure Logic Apps, or expose it as an API or MCP server?
Yes, three ways: call the REST API directly from a Logic App HTTP action; use a webhook start event for a dedicated, secured HTTP endpoint; or add the new MCP start event template, which exposes the process as a tool through the Processes MCP Server.

Can I interact with user tasks via MCP?
Yes. The Orchestration Cluster MCP Server now covers user tasks (search, get, assign, complete, read variables), plus incidents, process instances and definitions, using the same permissions as the REST API. 8.10 adds a second server, the Processes MCP Server, for your own deployed processes.

Is ProcessOS a separate license?
ProcessOS is in Early Access with 8.10, available to Enterprise customers and certified partners through the Download Center, for both SaaS and Self-Managed. It will be part of the Enterprise package once generally available. Talk to your account team to get started.

Will ProcessOS be fully in Camunda Hub for 8.11?
The ProcessOS UI is coming to Hub in 8.11, for both SaaS and Self-Managed. Where the ProcessOS worker actually runs is still open: fully in SaaS, or a hybrid setup where you run it yourself, which is likely needed when it has to reach local data. The final model will be documented closer to 8.11.

Any plan for reusable custom FEEL functions, the way connector templates are reusable?
Not in 8.10. Today you can package a FEEL expression inside a script task or a DMN-calling business rule task, save it as an element template, and share it through the Hub catalog: that's a reusable building block. Writing a function you call directly inside any FEEL expression is a different, not-yet-available capability. Camunda's product team is in early discovery of it and wants specific use cases.

Hub is the rollout starting point. What about worker code and its deployment blueprint?
Hub doesn't deploy or manage your workers in 8.10. Your CoE keeps ownership of worker and connector runtimes in your own repos and CI/CD. Hub covers modeling: your CoE stores approved element templates in Git, a pipeline submits them to the Hub catalog, and the catalog publishes only the template and its description. The implementation, job workers, BPMN, forms, stays in your repository and your own deployment pipeline.

Any plan for a BPMN-wide dictionary of values, to prefill variables, data objects, systems and roles consistently?
Camunda is tracking interest in this. Today, the closest option is organization-wide element templates shared through the Hub catalog: they can prefill variables, dropdowns and fixed properties so teams model with the same terms. Full administrator-defined JSON schemas across every BPMN element aren't available yet; share a specific use case and the product team will take it into the roadmap.

Watch the full session

The recording covers all of this live, including a full walkthrough of Elena Marku's KYC check end to end: a running agent in Operate, the decision trail compliance sees after, and the investigator's final approval.

Watch the recording

Camunda is building its product roadmap with input from the people who use it. If you want a say in what ships next, sign up for Camunda product research. Participation is voluntary, and you can opt out anytime.

Interested in ProcessOS? It's Enterprise-only during Early Access. Reach out to your account team to get started.

Start the discussion at forum.camunda.io

Ready to get started?