building-pydantic-ai-agents

This fragment documents workspace configuration and includes an agent tool that can execute model-supplied commands as host subprocesses. `LocalWorkspace` provides no isolation, so this is a significant capability that should be exposed only to trusted workflows with appropriate policy controls; the documentation warns against forwarding all environment variables because that could expose secrets such as LLM API keys. No evidence of malicious intent or covert data exfiltration appears in the fragment.