Agentwerke User Manual

Agentwerke is a governed software-delivery control plane for AI agents. It lets a team model delivery work as BPMN, assign steps to agents, gate tool use through policy, pause for human approvals, and keep an evidence trail for every run.

This site is the operator and developer manual for running Agentwerke, authoring workflows, configuring integrations, and extending the platform.

Choose your path

Start in five minutesRun the tokenless quickstart stack and complete the sample workflow.Operate a runStart a workflow, inspect its BPMN state, and understand step outcomes.Approve and auditHandle human gates and download an evidence pack for the run.Deploy AgentwerkeMove from quickstart to a hardened self-hosted deployment.Develop locallyBuild, test, and extend the .NET control plane and React UI.Use the referenceLook up BPMN extensions, policy boundaries, schema, and ADRs.

What you can do with Agentwerke

• Model SDLC workflows as versioned BPMN instead of prompt transcripts.
• Run model-backed or deterministic agent tasks with bounded permissions.
• Route every connector and tool call through the Tool Gateway.
• Pause at human approval tasks, including high-risk policy decisions.
• Run code-producing work inside Docker, OpenSandbox, or Kubernetes-backed sandboxes.
• Export evidence packs that include prompts, redactions, policy decisions, tool invocations, approvals, model usage, artifacts, and audit records.
• Keep the core platform self-hosted under your own data boundary.

Before production

The quickstart is intentionally easy: it uses a deterministic mock model provider and development authentication defaults. Production deployments must configure real authentication, secret handling, storage, sandbox isolation, and observability. Start with Deployment, then review Settings And Secrets and Auth And Data Residency.