Source: www.databricks.com/blog/introducing-unity-gateway-insights
Databricks says trace analysis of its own coding agent rollout turned up several ways to cut costs. Offline replays showed no performance gain from moving from Opus 4.8 to Opus 5, so the team stayed on the cheaper 4.8. Fixing common MCP errors recovered an estimated $1.2 million in costs a year, and traces showing simple prompts going to expensive models led it to adopt Smart Routing.
Insights is opened from AI Gateway > Observability > Insights and can generate reports covering the last day, 7 days, or 30 days. It compares each workstream's share of sessions with its share of tokens, ranks each kind of waste by cost, and links every finding to example sessions. Users can add up to 10 questions of their own and continue the investigation in Genie.
The feature relies on a complete record of each agent session, which a standalone gateway cannot provide. When engineers launch Claude Code, Codex, or GitHub Copilot with the ug CLI, Unity Gateway enables each agent's OpenTelemetry tracing automatically and writes traces to a governed Delta table in Unity Catalog. Only the table owner can query it until admins grant access, because traces contain prompts, code, and tool output.
Why it matters if you build agents
Teams running coding agents can use these reports to find failing MCP tools and frontier-model calls on simple tasks before the costs build up. Owners of MCP servers can trace a recurring error to a specific server and its owner. Because traces hold prompts and code, access to the trace table has to be granted deliberately.
Get stories like this every weekday
One short email with the day's agent tooling news.