Part IV — Production

The layered architecture behind an agentic application -- LLM, agent, tool, data, API, interface, observability -- shown as Agno's AgentOS: a REST API, a shared database, output_schema, and MCP-server exposure.

Separates deterministic regression testing of Part III's Python functions from ReliabilityEval-based evaluation of tool selection and trajectory, and builds an evaluation dataset targeting known ambiguities.

Tool-level validation versus agent-level guardrails (pre_hooks, InputCheckError): prompt injection, a custom finance-topic guardrail, memory and iteration limits, and approval checkpoints -- closing on defence in depth.

Agno's OpenTelemetry-based tracing, what to monitor -- trajectories, errors, latency, cost -- and deployment concerns: shared database, secrets, stateless workers, model-ID configuration.