This .NET 10 Microsoft Agent Framework sample reads the included mock holdings.csv through scoped host file tools, then uses HyperlightCodeActProvider to execute generated Python in a micro-VM.

Current policy: CodeActApprovalMode.NeverRequire. Generated sandbox code runs automatically, without a human approval prompt. The startup banner states that policy. Sandbox isolation is not human review or a general production security guarantee. Use only the included mock educational data; this is not financial advice.

From the public repository's session-03 directory:

dotnet run --project .\samples\31-codeact-agent\MafClaw.Sample31.csproj

Configure the existing Session 3 Foundry settings and Azure CLI identity first. Hyperlight needs hardware virtualization. On Windows, use dotnet run or the generated .exe apphost, not dotnet MafClaw.Sample31.dll.

Ask: What is the total portfolio value, and what percent is in Technology? The seeded four-row fixture totals $27,124.95, with 66.01% in Technology. The console prints actual TOOL CALL and TOOL RESULT messages separately from the assistant's answer. /exit closes the conversation.

The FileSystemAgentFileStore root remains the executable's working directory, and the existing read-only host-tool auto-approval rule remains in place. Any other approval requested by the framework is still presented to the console user. Unrelated mode-switching, file-memory, todo, skills, and web-search providers are disabled: this avoids a plan-first pause or automatic planning-note writes and keeps the lesson on file access plus CodeAct.

To demonstrate approval before each generated-code execution instead, change the visible assignment in Program.cs to CodeActApprovalMode.AlwaysRequire and rebuild. The frozen Session 3 slides describe that earlier policy; their approval narration is not the current default. See the current setup guide and demo script rather than inferring the policy from an older slide.