Agent Analytics built-in signals

Built-in signals are the quality checks Amplitude runs on every closed agent session, with no setup. Each signal answers one question about the session, such as whether the agent completed the task. Its result lands as a property on the session's [Agent] Session Record event, so you can chart, filter, and build cohorts on it right away.

Use built-in signals for questions every agent shares. For checks specific to your product, such as whether the agent quoted the right policy, add a custom evaluator instead. Refer to Set up custom evaluators.

Signals at a glance

Every signal writes a rationale: a short explanation of why the session got its result. Where a signal writes evidence, it points to the part of the session that supports the result. Signals are directional indicators, not ground truth. Read the rationale before you act on a single session, and trend signals across many sessions.

Task Completed

Did the agent complete the user's request by the end of the session? Task Completed credits recovery: a session where early tool calls fail but the agent then delivers what the user asked for counts as completed. It doesn't judge tone, formatting, or writing style.

Properties: [Agent] Task Completed, [Agent] Task Completed Rationale, [Agent] Task Completed Evidence.

Response Quality

Were the agent's responses accurate, clear, and well-structured? Response Quality judges the response text independent of whether the task got done, which Task Completed measures. A brief, clear answer that addresses the user counts as high quality.

Properties: [Agent] Response Quality, [Agent] Response Quality Rationale, [Agent] Response Quality Evidence.

Session Safety

Was the conversation legitimate usage, or adversarial or off-topic? Session Safety assigns one label per session.

Properties: [Agent] Session Safety, [Agent] Session Safety Rationale.

User Intent

What type of request was the user making? User Intent classifies the user's request, not the agent's response. When a conversation evolves, it uses the dominant intent across the session, and a greeting that leads into a task counts as the task.

Properties: [Agent] User Intent, [Agent] User Intent Rationale.

User Friction

Did the conversation show patterns that suggest a rough experience? Examples include the user repeating the same request, long clarification loops, consecutive tool failures, early abandonment, and the agent losing earlier context. User Friction reads behavior, so it can flag a session even when the user never complains.

Properties: [Agent] Has User Friction, [Agent] User Friction Rationale, and [Agent] Detected User Friction, which lists the patterns found.

Has Negative Feedback

Did the user explicitly express dissatisfaction with the agent? Examples include saying an answer is wrong, frustration aimed at the agent, rejecting an answer as a complaint, giving up or asking for a human, and a low rating. Normal iteration, such as rephrasing a request or asking a follow-up, doesn't count. Behavioral patterns without a complaint belong to User Friction.

A thumbs down or low rating that your app sends as an [Agent] Score also counts. Refer to Collect user feedback.

Properties: [Agent] Has Negative Feedback, [Agent] Negative Feedback Rationale, and [Agent] Detected Negative Feedback, which lists what the user said.

Data Quality Issues

Did the agent's responses or tool calls have mechanical problems? Response problems include empty responses, errors shown to the user, truncated output, repeated content, and refusals. Tool problems include failed or malformed calls, timeouts, rate limits, and provider errors.

Properties: [Agent] Has Data Quality Issues, [Agent] Data Quality Rationale, [Agent] Data Quality Evidence, and [Agent] Detected Data Quality Issues, which lists the issues found.

Signals and privacy modes

Most signals read message text. In metadata_only and customer_enriched, sessions carry no text, so those signal properties are absent from the Session Record rather than reported as failures. Data Quality Issues still reports in every mode. Refer to Agent Analytics privacy modes.

Where to analyze signals

Signal results appear on the Agent Analytics dashboards and on each session in the session viewer. To build your own charts, use the [Agent] Session Record event and its signal properties. Task Completed and Response Quality are true or false, so trend them as the share of sessions where the value is True. For the full Session Record property list, refer to Agent Analytics taxonomy.

Last verified on October 6, 2026.