How We Use Insights to Improve the Agent Behind Insights
Datadog, Friday, October 30th, 2026
Datadog describes using Insights in Agent Observability to find cost and reliability problems in its own AI agent.
Datadog's Insights analyzes agent traces to find recurring cost and reliability problems, explain causes, assess impact and recommend evidence-backed fixes, and Datadog applied it to its own agent.
One finding was that the search_llmobs_spans tool could return more content than needed: five search calls across two runs returned roughly 80,000 tokens that were resent to the model, with one result adding 19,000 tokens to a 97,000-token request.
Recommended fixes included narrower filters, smaller result pages and compact previews, which Datadog validates by replaying investigations and comparing tool output, token use and answer quality.
The post also covers investigating tool call failures.