Foundry Toolbox Autopilot

• Keep the sample focused on one Agent Framework agent consuming one dedicated, read-only Foundry Toolbox with anonymous Microsoft Learn and native Bing-backed Foundry Web Search. Only public search terms and URLs belong in tool arguments. Write actions, interactive consent, approval cards, workflows, and conversation history are out of scope.
• agent/app.py uses ActivityAgentServerHost(digital_worker=True) and its unmodified M365 stack. Preserve /activity/messages, the Teams selectors, and separate azd deploy / azd ai agent publish workflows.
• Create the agent and FoundryToolbox inside each Activity turn using agent/toolbox.py:create_agent. Never share MCP connections across turns: their background tasks inherit caller context at connection time. Preserve the SDK's platform headers for both model and Toolbox calls. Never synthesize user headers or call IDs.
• Use managed identity for hosted model and Toolbox access, and Azure CLI credentials for local execution. Neither source requires downstream user credentials, project connections, or secrets. Native Web Search is billed under Bing grounding terms, not a free anonymous external MCP service.
• Validate that TOOLBOX_ENDPOINT belongs to the configured Foundry project before sending a bearer token. Keep the positive allowlist in agent/toolbox.py consistent with toolbox.yaml; require both sources. Native Web Search is type: web_search, named web_search, with a required string search_query. Preserve source links in its embedded text-resource results. Distinguish official Microsoft guidance from other web sources.
• Learn wire names may use . or ___ between the server label and tool. Accept only the explicit aliases in TOOL_NAMES, reject duplicate aliases, and use exact discovered names for SDK filtering and calls. Native web_search is unprefixed; do not add an MCP server label or rename it.
• Preserve resource cleanup and explicit errors for empty tools, missing sources, unsupported approval policies, failed calls, and empty model replies.
• AgentServer owns telemetry initialization. Preserve the sensitive-content opt-out and disable only unused OpenAI Agents SDK instrumentation.
• Reuse the sibling guided-provisioning script and create/reuse Bicep pattern. Provisioning must not alter reused resources or role assignments. Manage the toolbox separately with azd ai toolbox create --from-file toolbox.yaml. Preserve its labels and read-only allowlists; copy azd's version-pinned MCP endpoint to TOOLBOX_ENDPOINT. When choosing a new version, refresh the endpoint from azd ai toolbox show --output json and redeploy.
• Use a capability-first README: short numbered setup steps, shared Hello World prerequisites and approval guidance, then example questions and expected Teams replies. Keep the sample-specific Toolbox requirements and provisioning/configuration/deployment commands here. Assume familiarity only with Hello World, not with other samples. Explain Agent Framework's value for this agent's tool handling without sample comparisons. Keep local development and offline tests out of the required deployment path. Focus the demonstration on questions and user-visible answers, not traces or SDK internals. Keep maintenance guidance brief; developer details belong here or in the source, not in a long post-demonstration appendix. Do not refer to development history or duplicate the full common setup guide.
• Keep requirements.txt fully pinned with cross-platform markers, including transitives. Do not make pywin32 unconditional. Use Python 3.13 explicitly; use a fresh virtual environment when changing Python minor versions.
• Run python -m unittest discover -s tests -v in the sample environment. Tests use real SDKs with HTTP test transports; preserve both-source execution, concurrent caller isolation, Activity ingress, cleanup, and failure coverage. Offline transports do not prove live Foundry permissions or hosted deployment.