FDE Times: Building and Evaluating Data Agents — grade the process, not just the answer
A 6 November FDE Times review of DeepLearning.AI and Snowflake’s ~2-hour Intermediate course says the FDE payoff is Agent GPA: an answer that is relevant and grounded can still be wrong if the planner queried the wrong quarter—so score goal, plan and action, not only the final text.
FDE Times on 6 November reviews DeepLearning.AI’s Building and Evaluating Data Agents course (1h59m, Intermediate, last updated 31 August 2026), taught by Snowflake’s Anupam Datta and Josh Reini. The longest lesson—"Measure Agent’s GPA" at 29 minutes—does not teach how to build an agent; it teaches how to check whether the agent did what it set out to do. For FDEs, a demo Q&A agent is easy; explaining Tuesday’s wrong answer and proving the fix is the hard part.
The course separates a planner from specialised data-fetching sub-agents, then uses LLM-as-a-judge for relevance and groundedness. Snowflake’s Agent GPA framework (goal–plan–action) catches the case where SQL filtered Q2 instead of Q3: the summary is still grounded in retrieved data and still wrong. Goal–action metrics cover correctness, relevance and groundedness; goal–plan covers plan quality and tool selection; plan–action covers adherence and tool calling. Inline evaluation lets the agent re-plan on low retrieval scores, trading latency for reliability—measure both configurations on real questions before production.
Treat platform-specific Snowflake bits as illustrations. The guide suggests taking the course, reading Snowflake’s GPA blog for the metric table, then citing arXiv 2510.08847 only when you need to defend the method to the customer’s engineering team—and putting a small public-data agent plus a goal/plan/action error table on your CV when postings mention evals.
Source: fdetimes.net