AWS Bedrock agent evaluation now works across all frameworks

The announcement addresses a pain point for teams building production agents: evaluation tooling historically required a specific SDK, LLM client, or tracing pattern. By standardizing on OpenTelemetry, Amazon Bedrock AgentCore Evaluations lets teams switch frameworks without losing the ability to score agent performance. The service reads three span roles—invoke agent, inference, and execute tool—from OpenTelemetry GenAI conventions or OpenInference, bridging both schemas. This means teams can use whatever framework suits their needs—LangGraph for workflow, LlamaIndex for retrieval, or OpenAI Agents SDK for GPT standardization—and still compare results with the same evaluators.

The coverage extends beyond the named frameworks via generic paths for OpenTelemetry or OpenInference instrumentation. Any library that deliberately conforms to these schemas is supported, making it easy for new frameworks to integrate. The main caveat is data grouping: the session.id must match, and message content must be present. For existing agents with pre-unified observability, this requires including the correlated event records; otherwise, response-quality evaluators fail. This is a practical limitation teams should plan for when adopting the service.