One finance-agent definition, multiple hosts, and explicit operational boundaries. This is a .NET 10 educational reference, not a production financial service or compliance certification. All portfolio/trade data is synthetic.

Download the Session 4 slides (PDF, 24 pages). The light .NET-purple comic presentation starts with the recap and audience follow-ups, then the operational lessons, and ends with the complete finance app.

The package contains real implementations, replacing the old .NET 9 placeholder:

• code: shared finance contracts/factory plus console, evaluation and Responses hosts.
• samples: seven plain-C# / MAF pairs, an Aspire observability variant, and two conditional service integrations.
• docs: setup, MCP tools, session persistence, Foundry Local, architecture, telemetry, policy, evaluations, deployment and operations.
• tests, evals, observability, scripts: executable evidence and operating assets.

Start with the recap, then these three focused topics before the Session 4 observability, governance, evaluation and hosting lessons:

These are independently runnable teaching samples, not additional capabilities of FinanceAgentFactory. The complete finance app composes the earlier session features and Session 4 operational controls at the end of the lesson. The sample ladder contains every run command.

For the live walkthrough: run .\tools\configure-user-secrets.ps1 once from the repository root, then enter a numbered sample folder and use dotnet run. No --live or save/resume arguments are needed. Samples 60/61 show the entire persistence flow in one run. Samples 70/71 use local models and need no cloud settings; pre-cache them before streaming. For Sample 12, first start aspire dashboard run in another terminal. Then dotnet run exports the generic agent's traces, metrics and structured logs to the local dashboard.

Every numbered sample is independent of the final app's projects and builds its own relevant MAF/Harness integration. The walkthroughs use tiny generic examples: a lesson-topic tool for observability, a mock outbox for approval, 2 + 3 for evaluation, and a greeting for hosting. MCP, persistence and local inference also have no financial-app dependencies. The cumulative financial application remains in code for the final reveal.

The published v04 slides predate these generic sample rewrites. Use the current source and guides for the revised demos; slide regeneration is deferred until the sample review is complete.

From this session directory:

dotnet run --project .\code\Console\MafClaw.Session04.Console.csproj -- --fixture
# Enter: Value the snapshot.
# Enter: /exit
dotnet run --project .\code\Evals\MafClaw.Session04.Evals.csproj -- --mode fixture
.\scripts\verify-session.ps1 -Mode Offline

Fixture mode uses scripted inference with actual MAF tool execution. It is not a live model demonstration. The evaluation corpus includes 24 deterministic contracts; additional regressions exercise approval, memory confinement, telemetry transport and a real local Responses request.

From the repository root:

.\tools\configure-user-secrets.ps1 -Session 4
.\tools\configure-user-secrets.ps1 -Session 4 -Check
dotnet run --project .\session-04\code\Console\MafClaw.Session04.Console.csproj

The full local profile adds files, a named local memory provider, skills, approval-gated shell/Hyperlight, and bounded research. Use --no-shell or --no-codeact explicitly if the corresponding local capability is unavailable. There is no silent fallback to unrestricted Python or fake inference.

The local implementation and its deterministic checks are separate from cloud certification. A local HTTP 200 is not enough: the Responses envelope must say completed. Hosted deployment, production identity/RBAC, Purview licensing and remote Foundry grading require their own configured, authorized verification. No cloud deployment is implied by this package.

See the sample ladder, setup, architecture, and runbook. The audience slides contain no presenter notes. Editable deck sources, presenter material and production evidence remain private.

• Session 4 event
• Series
• Series introduction
• Public repository
• Official production-ready article

Not financial advice. Do not put personal, confidential or real financial data in the model, tools, memory, logs or demonstrations.