View a markdown version of this page

Claude Agent SDK - Amazon Bedrock AgentCore
Instrument your agentHow spans are identifiedHow evaluation fields are extractedExample spans with event recordsExample spans without event records

Claude Agent SDK

This page explains how to instrument a Claude Agent SDK agent, how spans are identified, and how evaluation fields are extracted.

Topics

Instrument your agent

You can instrument a Claude Agent SDK agent with the OpenInference instrumentation library (openinference-instrumentation-claude-agent-sdk). This library emits telemetry under the scope name openinference.instrumentation.claude_agent_sdk, which Amazon Bedrock AgentCore Evaluations reads.

When your agent runs with the AWS Distro for OpenTelemetry (ADOT), such as on Amazon Bedrock AgentCore Runtime, you do not need to add explicit instrumentation code. Adding the instrumentation library to your project’s dependencies is enough. ADOT discovers it at startup and activates it automatically.

Add the instrumentation library to your dependencies.

Note

Use version 0.1.3 or later. This is the earliest version tested with the evaluation service.

requirements.txt:

openinference-instrumentation-claude-agent-sdk>=0.1.3

pyproject.toml:

[project]
dependencies = [
    "openinference-instrumentation-claude-agent-sdk>=0.1.3",
]
Note

Instrumentation is one step in setting up observability. To export telemetry for evaluation, complete the full setup in Set up observability.

How spans are identified

Claude Agent SDK is instrumented with the OpenInference convention, so AgentCore Evaluations classifies spans using the openinference.span.kind attribute.

Span type Identifying attribute

Invoke agent

openinference.span.kind = AGENT

Execute tool

openinference.span.kind = TOOL

The Claude Agent SDK emits only AGENT and TOOL spans; it does not emit separate inference (LLM) spans. The model metadata (model name, token usage) and the agent response are carried on the AGENT span itself.

How evaluation fields are extracted

The Claude Agent SDK produces clean, plain-text agent input and output, so the user prompt and agent response require no special parsing. Tool results, however, arrive as Anthropic content blocks in the form [{"type": "text", "text": "…​"}]. AgentCore Evaluations unwraps these blocks and concatenates their text.

The location of this content depends on how telemetry was collected. The identifying attribute (openinference.span.kind) is on the span in both cases. For more information, see Spans, event records, and telemetry signals.

From event records

When telemetry is split, AgentCore Evaluations reads content from the event record correlated to each span:

  • User prompt and agent response: from the invoke agent span’s event record, in body.input and body.output.

  • Tool call: the tool name from the tool.name attribute and the tool call ID from tool.id on the execute tool span. The tool arguments and result come from that span’s event record, in body.input and body.output. AgentCore Evaluations unwraps the Anthropic content blocks in the tool result.

For examples, see Example spans with event records.

From span attributes

When telemetry is not split, the same content stays on the span as attributes:

  • User prompt and agent response: from input.value and output.value on the invoke agent span.

  • Tool call: the tool name from tool.name, the tool call ID from tool.id, the arguments from input.value, and the result from output.value, on the execute tool span. AgentCore Evaluations unwraps the Anthropic content blocks in the tool result.

For examples, see Example spans without event records.

Example spans with event records

When telemetry is split, the span carries the identifying attributes and the content lives in a correlated event record. The following examples are from a Claude Agent SDK travel-planning agent deployed on Amazon Bedrock AgentCore Runtime.

Note

These examples are not complete spans. They show representative data from a real agent interaction, with some fields omitted and long values truncated for readability.

Example
Invoke agent span

The openinference.span.kind attribute (AGENT) identifies this as an invoke agent span. The span carries the model metadata; the conversation content lives in the correlated event record.

{
  "traceId": "6a292d74406894815807e2751e61dd49",
  "spanId": "a63aab3320ed8718",
  "name": "ClaudeAgentSDK.ClaudeSDKClient.receive_response",
  "kind": "INTERNAL",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk",
    "version": "0.1.5"
  },
  "attributes": {
    "openinference.span.kind": "AGENT",
    "llm.system": "anthropic",
    "llm.model_name": "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    "input.mime_type": "text/plain",
    "output.mime_type": "text/plain",
    "session.id": "sea-nyc-trip-2-turns-claude-adot"
  },
  "status": {
    "code": "OK"
  }
}
{
  "spanId": "a63aab3320ed8718",
  "traceId": "6a292d74406894815807e2751e61dd49",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk"
  },
  "body": {
    "input": {
      "messages": [
        { "role": "user", "content": "Hey, how can you help me" }
      ]
    },
    "output": {
      "messages": [
        { "role": "assistant", "content": "Hello! I'm your travel planning assistant ..." }
      ]
    }
  }
}
Execute tool span

The openinference.span.kind attribute (TOOL) identifies this as an execute tool span; tool.name holds the tool name and tool.id the tool call ID. The tool result lives in the correlated event record as Anthropic content blocks, which AgentCore Evaluations unwraps.

{
  "traceId": "6a292deb7450b3155895da4f38cb579a",
  "spanId": "909dcb4eb5f851ae",
  "name": "mcp__travel__search_flights",
  "kind": "INTERNAL",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk",
    "version": "0.1.5"
  },
  "attributes": {
    "openinference.span.kind": "TOOL",
    "tool.name": "mcp__travel__search_flights",
    "tool.id": "toolu_bdrk_01KmJhCRuEJJo6fswHbjCgFp",
    "tool.parameters": "{\"origin\": \"SEA\", \"destination\": \"NYC\", \"date\": \"2025-03-15\"}",
    "input.mime_type": "application/json",
    "output.mime_type": "application/json",
    "session.id": "sea-nyc-trip-2-turns-claude-adot"
  },
  "status": {
    "code": "OK"
  }
}
{
  "spanId": "909dcb4eb5f851ae",
  "traceId": "6a292deb7450b3155895da4f38cb579a",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk"
  },
  "body": {
    "input": {
      "messages": [
        { "role": "user", "content": "{\"origin\": \"SEA\", \"destination\": \"NYC\", \"date\": \"2025-03-15\"}" }
      ]
    },
    "output": {
      "messages": [
        {
          "role": "assistant",
          "content": "[{\"type\": \"text\", \"text\": \"{\\\"origin\\\": \\\"SEA\\\", \\\"destination\\\": \\\"NYC\\\", \\\"flights\\\": [ ... ]}\"}]"
        }
      ]
    }
  }
}

Example spans without event records

When telemetry is not split, the same content stays on the span attributes and no separate event record is produced. The following examples are from a Claude Agent SDK travel-planning agent.

Note

These examples are not complete spans. They show representative data from a real agent interaction, with some fields omitted and long values truncated for readability.

Example
Invoke agent span

The input.value attribute holds the user prompt, and the output.value attribute holds the agent response, both as plain text.

{
  "traceId": "561876bb17e9eaeb2f194ee515742b2f",
  "spanId": "3b6815f5b3909a51",
  "name": "ClaudeAgentSDK.ClaudeSDKClient.receive_response",
  "kind": "INTERNAL",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk",
    "version": "0.1.3"
  },
  "attributes": {
    "openinference.span.kind": "AGENT",
    "llm.system": "anthropic",
    "llm.model_name": "us.anthropic.claude-sonnet-4-5-20250929-v1:0",
    "input.value": "Hey, how can you help me",
    "input.mime_type": "text/plain",
    "output.value": "Hi there! ... How can I help you plan your next adventure?",
    "output.mime_type": "text/plain",
    "session.id": "sea-nyc-trip-2-turns-claude-unified"
  },
  "status": {
    "code": "OK"
  }
}
Execute tool span

The input.value attribute holds the tool arguments, and the output.value attribute holds the tool result as Anthropic content blocks, which AgentCore Evaluations unwraps.

{
  "traceId": "7bb7e59a30d03fc0b9da5bf009a3b429",
  "spanId": "d27b488965bbba99",
  "name": "mcp__travel__search_flights",
  "kind": "INTERNAL",
  "scope": {
    "name": "openinference.instrumentation.claude_agent_sdk",
    "version": "0.1.3"
  },
  "attributes": {
    "openinference.span.kind": "TOOL",
    "tool.name": "mcp__travel__search_flights",
    "tool.id": "toolu_bdrk_019yE7Gne1rZKE3UnVPAWjLq",
    "tool.parameters": "{\"origin\": \"SEA\", \"destination\": \"NYC\", \"date\": \"2025-03-15\"}",
    "input.value": "{\"origin\": \"SEA\", \"destination\": \"NYC\", \"date\": \"2025-03-15\"}",
    "input.mime_type": "application/json",
    "output.value": "[{\"type\": \"text\", \"text\": \"{\\\"origin\\\": \\\"SEA\\\", \\\"destination\\\": \\\"NYC\\\", \\\"flights\\\": [ ... ]}\"}]",
    "output.mime_type": "application/json",
    "session.id": "sea-nyc-trip-2-turns-claude-unified"
  },
  "status": {
    "code": "OK"
  }
}