Observe AI agents that run on AWS Bedrock.
Your agent calls Bedrock — Claude, Nova, Titan, Mistral, Llama — for inference. Trefur captures each invocation as a span on the agent trace with tokens, latency, cost, and any tool-use cycle that came with it.
What you see.
Why it matters.
For engineers
Drop in the Trefur SDK. Existing boto3 calls to Bedrock are auto-instrumented. Every InvokeModel and Converse call shows up on the agent trace with tokens, latency, and cost.
For platform teams
Roll up Bedrock usage per account, per region, per model ID, per agent. Give finance a clean breakdown without joining CloudWatch logs to product telemetry by hand.
For business operators
Know which workflows depend on Bedrock and what they cost. Catch silent failover to a more expensive model. Tie model spend to ticket resolution or completion rate.
For SRE
Bedrock throttling and capacity errors land on the span as tool failures. Envelope alerts catch retry-rate spikes and latency walks before they reach customers.
How it works.
Your agent calls Bedrock through the AWS SDK (boto3, AWS SDK for JavaScript, etc.). The Trefur SDK wraps the Bedrock client and emits a span per InvokeModel or Converse call. Spans use the OpenTelemetry GenAI shape so they land in the same trace as any other LLM call from a different vendor. The collector runs inside your network if you want raw prompt content to stay local — egress is configurable.
See what your Bedrock agents are actually doing.
Free tier. No card required. First trace in under five minutes.