Seven-part hands-on tutorial
Build agents that enterprises can trust¶
Design a governed Java agent platform one control at a time—with MCP, A2A, Quarkus, agentgateway, OpenTelemetry, human approval, and continuous evaluation. Extend it with model routing using Agent Router and bounded code execution using OpenSandbox.
7guided tutorials
Java 25modern Quarkus stack
MCP + A2Aopen agent protocols
Local firstinspect every control
Choose your path¶
01
Learn the platform
Follow the seven-part path from typed MCP tools through model routing and governed sandbox execution.
▶
Run the demos
Launch the local consoles and model-routing terminal lab. No hosted LLM is required for the guided flows.
◇
Plan for production
Explore identity, policy lifecycle, durable approvals, telemetry privacy, and failure modes.
A realistic governance scenario¶
The examples follow Acme FinServ, a fictional B2B payments company operating under SOC 2, PCI-DSS, and GDPR obligations. The code is intentionally local and inspectable; every tutorial separates the control demonstrated in the lab from the work required for production.
The control chain¶
| Part | Control | Evidence produced |
|---|---|---|
| 1 | Typed schemas and boundary validation | Rejected malformed tool arguments |
| 2 | Agent authentication, tool authorization, guardrails | Identity and policy decisions |
| 3 | Trace propagation and telemetry | End-to-end request timeline |
| 4 | Workflow policy, human approval, and Agent Skills | State transitions, approval record, and investigation evidence |
| 5 | Golden datasets and regression gates | Repeatable control test report |
| 6 | Model routing and bounded provider failover | Backend attempt counts and routing smoke checks |
| 7 | Bounded sandbox execution and cleanup | Reconciliation findings, failed egress probes, execution results, and deletion checks |
Lab scope
These examples teach control mechanics. They are not a compliance certification, production identity provider, durable workflow engine, or complete threat model.