Key Benefits

• Structured interpretation – Converts natural-language database requests into a normalized change_json contract.
• Separation of responsibilities – Uses specialized agents for parse, impact analysis, rollout planning, and SQL generation.
• Optional RAG grounding – Lets the impact-analysis agent use File Search over an uploaded PDF, such as an IFD, schema spec, or lineage document.
• Typed stage outputs – Uses Pydantic models and Agents SDK output schemas for parse, impact, and plan stages.
• Guardrail-first workflow – Adds deterministic checks between stages so obvious failures are caught before downstream steps consume bad state.
• Traceability – Emits OpenAI Agents SDK traces and spans for agent runs, guardrails, artifact generation, and eval execution.
• Portable artifacts – Saves the final workflow bundle as JSON under artifacts/notebook_runs/.
• Eval-ready design – Generates Promptfoo provider, assertion, config, and result files from the live notebook state.
• No database side effects – Produces draft SQL and validation output without executing against a live database.