View a markdown version of this page

Agent sessions - Amazon CloudWatch

Agent sessions

The Agent sessions view groups the traces that belong to the same conversation or user interaction, so you can follow a multi-turn exchange from start to finish and see where it went wrong across turns. For how traces, spans, and sessions relate, see Monitor AI agents.

Why sessions need their own view

Some failures are invisible at the single-trace level because every individual turn looks fine. The Agent sessions view surfaces the cross-turn patterns:

  • Lost context — the agent asks for information the user already provided, or contradicts an earlier answer.

  • Churn — the same tool is called with the same arguments across consecutive turns, or the user rephrases the same request repeatedly (a sign the agent is not making progress).

  • Degradation over length — quality that is fine on turn one and poor by turn eight, which single-trace sampling never shows.

To follow a session

  1. Open your space, then choose Agent sessions.

  2. Select a session to see its turns in order and read the conversation flow.

  3. Select the turn where the problem starts to open its trace in Agent traces.

A session in Agent sessions showing its turns in order with summary cards, where a later turn surfaces a problem that no single turn shows on its own.

Evaluate and reuse sessions

Cross-turn quality checks run from this view. Select one or more sessions, and then choose an action from the toolbar:

  • Evaluate — score the selected sessions. Evaluation runs through Amazon Bedrock AgentCore as a background batch job, and scores appear alongside the sessions as they complete. Only evaluators that score whole sessions can run here; evaluators that score individual traces run from Agent traces. For the evaluator catalog and how scoring works, see Evaluators and evaluations.

  • Create dataset or Add to dataset — turn the selected sessions into test cases, so a multi-turn scenario becomes a repeatable test. You choose whether each session becomes one test case or each of its traces does. See Datasets.

Group traces into sessions

Omni groups traces into sessions using the OpenTelemetry session.id span attribute. Agents hosted on Amazon Bedrock AgentCore, and agents built with frameworks that set the attribute natively (such as Strands Agents), are grouped automatically. Custom and non-framework agents must set the attribute themselves. Otherwise, each turn appears as an unrelated trace.

Set session.id on the root span of every turn that belongs to the same conversation, reusing one stable ID per conversation and starting a fresh ID when a new conversation starts:

from opentelemetry import trace

# Reuse the SAME id for every turn in one conversation.
session_id = "conv-8f2a1c"  # your stable per-conversation id

span = trace.get_current_span()
span.set_attribute("session.id", session_id)

Note Only session.id groups traces into sessions. Setting the gen_ai.conversation.id attribute is harmless (other tools read it), but Omni does not use it for session grouping.