Agentic tool calling
What is an agentic LLM with tool calling? How do you keep it safe in production?
Answers use simple, clear English.
Quick interview answer
The model proposes tool calls (JSON schema); an orchestrator executes tools and returns results until a final answer (ReAct-style loop). Safety: allowlisted tools, least privilege, timeouts/max steps, human approval for irreversible actions, sandboxes, audit logs, and treating retrieved/tool text as untrusted. Eval task success—not just fluent chat.
Detailed answer
The model proposes tool calls (JSON schema); an orchestrator executes tools and returns results until a final answer (ReAct-style loop). Safety: allowlisted tools, least privilege, timeouts/max steps, human approval for irreversible actions, sandboxes, audit logs, and treating retrieved/tool text as untrusted. Eval task success—not just fluent chat.
Real example & use case
Ops agent: read-only Kubernetes get; delete requires human confirmation.
Pros & cons
Pros: real actions beyond text. Cons: cost, loops, security risk if unconstrained.