TraceAI Strands Agents Integration

Comprehensive observability for AWS Strands Agents with TraceAI.

Strands Agents is an open-source SDK from AWS that enables building AI agents with a model-driven approach. This integration provides seamless tracing by leveraging Strands' built-in OpenTelemetry support.

Features

• Zero-config integration: Works with Strands' native OTEL support
• Automatic tracing: Agent invocations, tool calls, and model interactions
• Token tracking: Input/output tokens and cache metrics (Bedrock)
• Session correlation: Link traces across conversations
• Custom callbacks: Extended event capture for detailed observability
• MCP support: Trace Model Context Protocol tool usage

Installation

pip install traceai-strands

For full functionality with Strands:

pip install 'strands-agents[otel]'

Quick Start

Option 1: Using TraceAI with fi_instrumentation

from fi_instrumentation import register
from fi_instrumentation.fi_types import ProjectType
from traceai_strands import configure_strands_tracing

# Setup TraceAI
trace_provider = register(
    project_type=ProjectType.OBSERVE,
    project_name="my-strands-agent",
)

# Configure Strands to use TraceAI
configure_strands_tracing(tracer_provider=trace_provider)

# Now use Strands normally - traces are sent automatically
from strands import Agent

agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    system_prompt="You are a helpful assistant.",
)

response = agent("Hello!")

Option 2: Direct OTLP Configuration

from traceai_strands import configure_strands_tracing

# Configure with OTLP endpoint directly
configure_strands_tracing(
    otlp_endpoint="https://api.traceai.com/v1/traces",
    otlp_headers={"Authorization": "Bearer YOUR_API_KEY"},
    project_name="my-strands-agent",
)

from strands import Agent

agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    system_prompt="You are a helpful assistant.",
)

response = agent("Hello!")

Adding Trace Attributes

Add session and user information for better trace correlation:

from strands import Agent

agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    system_prompt="You are a helpful assistant.",
    trace_attributes={
        "session.id": "user-session-123",
        "user.id": "user@example.com",
        "tags": ["production", "chatbot"],
    },
)

Or use the helper function:

from traceai_strands import create_traced_agent

agent = create_traced_agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    system_prompt="You are a helpful assistant.",
    session_id="user-session-123",
    user_id="user@example.com",
    tags=["production", "chatbot"],
)

Using Tools

Strands tools are automatically traced when using the @tool decorator:

from strands import Agent, tool
from typing import Annotated

@tool
def get_weather(city: Annotated[str, "City name"]) -> str:
    """Get the current weather for a city."""
    return f"Weather in {city}: 72°F, Sunny"

@tool
def calculate(
    operation: Annotated[str, "add, subtract, multiply, divide"],
    a: Annotated[float, "First number"],
    b: Annotated[float, "Second number"],
) -> float:
    """Perform a calculation."""
    ops = {"add": lambda x, y: x + y, "multiply": lambda x, y: x * y}
    return ops[operation](a, b)

agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    tools=[get_weather, calculate],
)

response = agent("What's 15 times 7, and what's the weather in Tokyo?")

Custom Callback Handler

For extended event capture beyond Strands' built-in telemetry:

from traceai_strands import StrandsCallbackHandler

# Create callback handler
callback = StrandsCallbackHandler(
    tracer_provider=trace_provider,
    capture_input=True,
    capture_output=True,
)

# Use with agent
agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    callback_handler=callback,
)

# Lifecycle events are automatically traced
response = agent("Hello!")

MCP Integration

Trace Model Context Protocol server tools:

from strands import Agent
from strands.tools.mcp import MCPClient

# Connect to MCP server
mcp_client = MCPClient(
    server_command=["npx", "@anthropic/mcp-server-calculator"],
)

# Get MCP tools
mcp_tools = mcp_client.list_tools_sync()

# Create agent with MCP tools
agent = Agent(
    model="us.anthropic.claude-sonnet-4-20250514-v1:0",
    tools=mcp_tools,
)

response = agent("Calculate the square root of 144")

Semantic Attributes

The integration captures these OpenTelemetry GenAI semantic attributes:

Agent Attributes

Attribute	Description
agent.type	Agent class name
strands.system_prompt	Agent's system prompt (truncated)
strands.tool_count	Number of tools available
strands.session.id	Session identifier
strands.user.id	User identifier

Model Attributes

Attribute	Description
gen_ai.system	Model provider (bedrock, openai, etc.)
gen_ai.request.model	Model name/ID
gen_ai.request.temperature	Temperature setting
gen_ai.request.max_tokens	Max tokens setting
gen_ai.usage.input_tokens	Input token count
gen_ai.usage.output_tokens	Output token count
strands.cache.read_tokens	Cache read tokens (Bedrock)
strands.cache.write_tokens	Cache write tokens (Bedrock)

Tool Attributes

Attribute	Description
gen_ai.tool.name	Tool function name
gen_ai.tool.description	Tool docstring
gen_ai.tool.parameters	Tool input parameters
gen_ai.tool.result	Tool execution result

Environment Variables

Strands respects standard OpenTelemetry environment variables:

# OTLP endpoint
export OTEL_EXPORTER_OTLP_ENDPOINT="https://api.traceai.com/v1/traces"

# OTLP headers (authentication)
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer YOUR_API_KEY"

# Sampling (optional)
export OTEL_TRACES_SAMPLER="traceidratio"
export OTEL_TRACES_SAMPLER_ARG="0.5"  # Sample 50%

Model Provider Support

The integration automatically detects model providers:

Provider	Model Patterns
Bedrock	us.anthropic.*, us.amazon.*, eu.*
OpenAI	gpt-*, o1-*, text-davinci-*
Anthropic	claude-*
Google	gemini-*, palm-*
Mistral	mistral-*, mixtral-*
Meta	llama-*
Ollama	ollama/*

Examples

See the examples/ directory for complete examples:

• basic_agent.py - Simple agent with tracing
• agent_with_tools.py - Tools and function calling
• callback_handler.py - Custom callback handler
• mcp_agent.py - MCP server integration

How It Works

Strands has native OpenTelemetry support through its StrandsTelemetry class. This integration:

1. Configures OTLP export: Sets environment variables or configures StrandsTelemetry to send traces to TraceAI
2. Adds helper functions: Provides create_traced_agent() and attribute helpers
3. Optional callbacks: StrandsCallbackHandler for extended event capture

The integration is lightweight because Strands already does the heavy lifting for telemetry.

Compatibility

• Strands Agents: >= 1.0.0
• Python: >= 3.10
• OpenTelemetry: >= 1.0.0

Resources

• Strands Agents Documentation
• Strands GitHub
• OpenTelemetry GenAI Conventions
• TraceAI Documentation

License

Apache-2.0