Bedrock Agents

Trace Amazon Bedrock Agents invocations with Weave, capturing agent inputs, foundation model usage, and completion output.

Amazon Bedrock Agents let you build managed AI agents on AWS that orchestrate foundation models, knowledge bases, and action groups. Weave traces calls to the bedrock-agent-runtime client so you can inspect each invoke_agent invocation, including the foundation model, token usage, session ID, and the agent’s response.

The Weave TypeScript SDK doesn’t currently Bedrock Agent integration.

Prerequisites

• A W&B API key. For more information, see API keys.

• AWS credentials configured for an account with access to Bedrock Agents (see Authentication).

• An existing Bedrock agent and alias. Note the agentId and agentAliasId. You can return a list of agents associated with your ID by running the following AWS CLI command, replacing <your-region> with the region slug your agent resides in:

  aws bedrock-agent list-agents --region us-east-1
  

Installation

Install Weave and the AWS SDK:

pip install weave boto3

Trace invoke_agent calls

Create a bedrock-agent-runtime client and pass it to patch_client. Weave detects the client type and wraps the invoke_agent method. After patching, use the client as you normally would.

import boto3

import weave
from weave.integrations.bedrock import patch_client

weave.init("your-team-name/bedrock-agents-demo")

# Create and patch the Bedrock Agents runtime client.
bedrock_agent_client = boto3.client("bedrock-agent-runtime", region_name="us-east-1")
patch_client(bedrock_agent_client)

# Invoke the agent. Set `enableTrace=True` so Weave can capture the underlying
# foundation model and token usage from the orchestration trace events.
response = bedrock_agent_client.invoke_agent(
    agentId="[YOUR-AGENT-ID]",
    agentAliasId="[YOUR-AGENT-ALIAS-ID]",
    sessionId="[YOUR-SESSION-ID]",
    inputText="What is the capital of France?",
    enableTrace=True,
)

# Consume the streaming completion to assemble the final response text.
final_text = ""
for event in response["completion"]:
    chunk = event.get("chunk")
    if chunk and "bytes" in chunk:
        final_text += chunk["bytes"].decode("utf-8")

print(final_text)

Each invoke_agent call appears in the Weave UI as a BedrockAgentRuntime.invoke_agent trace. The trace records:

• The agent inputs (agentId, agentAliasId, sessionId, inputText).
• The extracted assistant text from the completion event stream.
• The underlying foundation model used by the agent (extracted from the orchestration trace).
• Token usage (prompt_tokens, completion_tokens, total_tokens) when reported by the agent.

Foundation model and token usage are extracted from orchestrationTrace events, which Bedrock only emits when invoke_agent is called with enableTrace=True. Without this flag, traces still capture the inputs and the generated response text, but the foundation model falls back to bedrock-agent:<agentId> and token counts are unavailable.

Wrap an agent call with weave.op

To group an invoke_agent call with related logic, such as preprocessing, postprocessing, or chained API calls, wrap your function with @weave.op. Weave nests the patched invoke_agent trace inside the parent op.

@weave.op
def ask_agent(question: str, session_id: str) -> str:
    response = bedrock_agent_client.invoke_agent(
        agentId="[YOUR-AGENT-ID]",
        agentAliasId="[YOUR-AGENT-ALIAS-ID]",
        sessionId=session_id,
        inputText=question,
        enableTrace=True,
    )
    text = ""
    for event in response["completion"]:
        chunk = event.get("chunk")
        if chunk and "bytes" in chunk:
            text += chunk["bytes"].decode("utf-8")
    return text


answer = ask_agent("Summarize today's open support tickets.", session_id="session-1")

Multi-turn conversations

Bedrock Agents preserve conversation state on the service side when you reuse a sessionId. To group multiple turns into a single trace in the Weave UI, wrap the turns in weave.thread:

with weave.thread("support-conversation") as t:
    for prompt in [
        "I can't log in to my account.",
        "I already tried resetting my password.",
    ]:
        ask_agent(prompt, session_id=t.thread_id)

Weave displays each turn in the UI as individual rows in the Threads view.

View traces

When you run the example, Weave prints a link to the project dashboard. Open the link to inspect the agent inputs, foundation model, token usage, and the generated response for each invoke_agent call.