Observability

AgentCore automatically enables:

• X-Ray tracing — every invocation generates a trace
• CloudWatch logging — agent logs ship to CloudWatch

These are on by default whether you’re running deployed (agentcore deploy + invoke) or locally (agentcore dev). The dev server auto-instruments your agent with the AWS OpenTelemetry distro the same way the deployed runtime does; opt out with agentcore dev --no-traces.

Two prerequisites for the local path to work end-to-end:

1. AWS credentials available locally — the OTEL exporter needs them to ship spans.
2. CloudWatch Transaction Search is enabled on the account (one-time setup per account) — see “Viewing traces” below. Without it, spans are ingested but not searchable, so agentcore traces list and agentcore run eval --session-id return empty.

After deploy, AgentCore Runtime also auto-instruments the container (the default CMD wraps the app with opentelemetry-instrument). You don’t need to configure OTEL in your code for either path — but you do need your agent code to be instrumented correctly.

Ensuring logs appear in CloudWatch

Three things must be true for logs to appear:

1. OTEL entrypoint wrapper in Dockerfile

Your Dockerfile CMD must use the OpenTelemetry wrapper:

For CodeZip builds, this is handled automatically. For Container builds, you must add it.

2. IAM permissions for CloudWatch and X-Ray

Your runtime execution role needs:

3. Use the logging module, not print()

AgentCore captures structured logs via the Python logging module. print() statements go to stdout but are not captured by the OTEL pipeline.

Viewing traces

Traces show the full execution path of one agent invocation — model calls, tool calls, and timing.

Trace delay: Traces appear ~10 seconds after invocation (previously 30–60s). Don’t panic if they’re not immediate, and don’t bake longer waits into scripts — older skills and docs that say “30–60 seconds” or “2–5 minutes” are stale.

Also verify Transaction Search is enabled in CloudWatch — this is a prerequisite for trace visibility in the console.

Viewing logs

CloudWatch dashboard

For production agents, set up a CloudWatch dashboard with:

• Invocation count and error rate
• P50/P90/P99 latency
• Memory and CPU utilization
• Error log count by type

These metrics are available in the AWS/BedrockAgentCore namespace after deploy.

Multi-account observability

If your agents are spread across accounts (typical setup: separate prod / staging / dev accounts), use CloudWatch cross-account observability to view metrics, traces, and logs from one central monitoring account.

The setup order matters — do it in this sequence or the console won’t show source-account data:

1. Pick a monitoring account. This is where you’ll view everything. Often a central observability account, not a workload account.
2. Configure the monitoring account first. CloudWatch console → Settings → Monitoring account configuration → Configure. Choose which telemetry types to share (enable Metrics and Logs — traces go through X-Ray’s own cross-account mechanism).
3. Link each source account. Either via AWS Organizations (if your accounts are in one) or via individual linking. Source accounts must accept the link.
4. Deploy AgentCore agents in the source accounts with observability enabled — same default OTEL wrap-up as single-account. No code changes needed.
5. View from the monitoring account. AgentCore Observability in the CloudWatch console now shows data from all linked accounts side-by-side, identified by source account ID.

Order-of-operations trap: if you deploy agents in source accounts before linking, the telemetry still flows correctly — it just won’t be visible from the monitoring account until the link is active. You don’t need to redeploy, just wait a few minutes after linking.

Traces: cross-account trace viewing uses X-Ray’s existing cross-account sharing model. If the CloudWatch cross-account link is set up correctly for Logs and Metrics but traces don’t show, check X-Ray’s cross-account config separately.

IAM: no extra IAM on the agent execution roles for cross-account observability. The cross-account feature operates at the CloudWatch/X-Ray layer, not at the source of the telemetry.

See cross-account observability (opens in a new tab) for the current console flow and edge cases.

Cross-references

• If logs aren’t appearing at all, check the three requirements above or use agents-debug
• For production observability setup, see agents-harden
• For measuring agent quality (not just operational health), load references/evals.md