Trace Signals: your agent finished. Did it actually help?
Find frustration, repeated requests, and unanswered questions in your agent traces. Cloudflare's specialized models make broad trace analysis affordable.
A completed request can still leave a user repeating themselves, waiting on a promised follow-up, or asking a question your agent never answers. Error rates won't tell you that.
Trace Signals helps you find those conversations without reading every trace yourself. And thanks to Cloudflare's new specialized models, cost doesn't have to force you to settle for a small sample.
Clairic extracts typed values. Clef Flash handles boolean and choice questions. Neither generates output tokens. That changes the economics of automatic analysis and moves us toward our goal: analyzing every trace, for every customer, without breaking the bank.
Start with the conversation_health preset to flag frustration, repeated requests, promised follow-ups, and unanswered questions. Results include values, confidence, and source excerpts. Filter your traces by those signals to find the conversations worth investigating.
Enable a signal set on an agent to analyze its eligible completed turns by default. Analysis runs asynchronously, outside the response path, and supports both Runtype agents and external agents sending OTLP traces. You can narrow coverage with filters or sampling; a monthly budget caps spending by skipping further analysis when reached.
Find the conversations that need attention, even when the agent reports success.