A Databricks agent gives accurate answers at the start of a session but increasingly cites irrelevant earlier tool output as the session grows. No hard token limit has been reached. What should the engineer measure first?

1. Only the number of rows in the underlying catalog.
2. Only whether the endpoint returns HTTP 200.
3. Context composition and quality by turn, including which earlier results remain in each model input.
4. Only the maximum context length advertised in the model card.