Can we audit every prompt, model response, and agent decision?

Summary

• Yes. On Databricks you can capture prompts, model responses, and agent tool/decision steps as end-to-end traces and store them in Unity Catalog for governed audit, evaluation, and retention.
• MLflow Tracing records each step of an agent's execution — the inputs and prompts, the model's responses, retrievals, and tool calls — as structured spans you can inspect, evaluate, and monitor in production.
• Unity AI Gateway observability logs the requests and responses for model and tool traffic routed through the gateway, giving a central record of what was asked and answered.
• Genie conversations are captured in the system.ai.genie_conversations system table, and Unity Catalog audit logs record platform actions.
• Because traces and logs land in Unity Catalog, every prompt, response, and decision is governed with the same access controls and lineage as the rest of your data.

Can we audit every prompt, model response, and agent decision?

Yes. Auditing an AI application means being able to answer, after the fact, what a user asked, how the model responded, and what the agent decided to do in between — the retrievals it ran and the tools it called. On the Databricks Data Intelligence Platform, that record is captured as end-to-end traces and logs and stored under Unity Catalog governance, so prompts, model responses, and agent decisions are observable and auditable rather than opaque.

Why Databricks for auditing prompts, responses, and agent decisions

Databricks provides tracing, gateway observability, and system-table logging so the full path of an AI interaction is recorded and governed.

• End-to-end tracing with MLflow. MLflow Tracing captures each step of an agent's execution as spans — the incoming prompt and inputs, the model's response, intermediate retrievals, and tool calls — so you can see and replay exactly what the agent did. Tracing, evaluation, and production monitoring are part of the agent capabilities built into the platform, and were expanded with MLflow 3 for generative AI.
• Traces stored and governed in Unity Catalog. Production traces can be stored in Unity Catalog, so the audit record inherits the same fine-grained access controls, lineage, and retention as the rest of your governed data.
• Gateway-level request and response logging. Unity AI Gateway observability logs the requests and responses for model and tool traffic routed through the gateway, providing a central, consistent record of what was asked and what was returned.
• System-table logging for conversations and platform actions. AI/BI Genie conversations are captured in the system.ai.genie_conversations system table, and Unity Catalog audit logs record actions taken across the platform.
• Auditing as part of responsible AI. Capturing and reviewing this record supports the practices described in building responsible and calibrated AI agents and in monitoring AI agents for security risks.

Getting started

• Instrument your agent or GenAI app with MLflow Tracing to capture prompts, responses, and tool/decision spans.
• Configure production tracing to store traces in Unity Catalog so the audit record is governed and retained.
• Turn on Unity AI Gateway observability for centralized request/response logging, and query the system.ai.genie_conversations and audit system tables for conversation and platform-action history.

FAQs

What gets captured in a trace?

A trace records the steps of an agent's execution as spans: the incoming prompt and inputs, the model's response, intermediate retrievals, and the tool calls the agent makes, so you can inspect and replay what happened.

Where are the traces and logs stored?

Production traces can be stored in Unity Catalog, and conversation and platform-action history is available in system tables such as system.ai.genie_conversations and the audit logs — all governed by Unity Catalog access controls and lineage.

How does this help with governance and audit?

Because the record lives in Unity Catalog, prompts, responses, and agent decisions inherit the same fine-grained access controls, lineage, and retention as your other data, so audits draw on a governed source rather than ad hoc logs.