feat(bedrock): trace bedrock runtime calls

Add a Python Bedrock Runtime integration that auto-instruments boto3/botocore client creation. The public auto instrumentation API is:

```python
import braintrust

braintrust.auto_instrument(bedrock=True)
```

The integration requires botocore >= 1.34.116, the first supported version whose Bedrock Runtime service model exposes both Converse and ConverseStream. It traces Converse, ConverseStream, InvokeModel, and InvokeModelWithResponseStream with Bedrock-shaped LLM spans.

Manual wrapping is also exposed for existing clients:

```python
from braintrust.integrations.bedrock_runtime import wrap_bedrock

client = wrap_bedrock(client)
```

Author: AbhiPrasad

examples/README.md

@@ -62,6 +62,7 @@ Unless noted otherwise, every example below uses `braintrust.auto_instrument()`.
 | `agno/` | Agno agents and teams, sync + async, streaming + non-streaming |
 | `anthropic/` | Sync and async Anthropic clients |
 | `autogen/` | AutoGen `AssistantAgent` backed by `OpenAIChatCompletionClient` |
+| `bedrock_runtime/` | boto3 Bedrock Runtime `converse()` call against Amazon Nova Lite |
 | `claude_agent_sdk/` | Claude Agent SDK subprocess query |
 | `cohere/` | Cohere `ClientV2` chat call |
 | `crewai/` | CrewAI `Agent` + `Task` + `Crew` end-to-end |

examples/bedrock_runtime/README.md

@@ -0,0 +1,25 @@
+# AWS Bedrock + Braintrust
+
+Calls `braintrust.auto_instrument()` to patch boto3's Bedrock Runtime client factory, then runs a single `client.converse(...)` call against Amazon Nova Lite. The trace captures a parent `answer_question` span plus a child `bedrock.converse` LLM span with input, output, metadata, and token metrics.
+
+## Run
+
+```bash
+export BRAINTRUST_API_KEY=...
+export AWS_BEARER_TOKEN_BEDROCK=...
+export AWS_REGION=us-east-1
+export AWS_DEFAULT_REGION=us-east-1
+
+uv sync
+uv run python example.py
+```
+
+If you use IAM credentials instead of a Bedrock API key, set the usual AWS variables instead:
+
+```bash
+export AWS_ACCESS_KEY_ID=...
+export AWS_SECRET_ACCESS_KEY=...
+export AWS_SESSION_TOKEN=... # if using temporary credentials
+```
+
+The default model is `us.amazon.nova-lite-v1:0`. Override it with `BRAINTRUST_BEDROCK_MODEL` if needed.

examples/bedrock_runtime/example.py

@@ -0,0 +1,38 @@
+#!/usr/bin/env python
+"""boto3 Bedrock Runtime client traced via braintrust.auto_instrument()."""
+
+import os
+
+import braintrust
+
+
+braintrust.auto_instrument()
+braintrust.init_logger(project="example-bedrock")
+
+import boto3  # noqa: E402
+
+
+MODEL = os.getenv("BRAINTRUST_BEDROCK_MODEL", "us.amazon.nova-lite-v1:0")
+REGION = os.getenv("AWS_REGION") or os.getenv("AWS_DEFAULT_REGION") or "us-east-1"
+
+
+@braintrust.traced
+def answer_question(question: str) -> str:
+    client = boto3.client("bedrock-runtime", region_name=REGION)
+    response = client.converse(
+        modelId=MODEL,
+        messages=[{"role": "user", "content": [{"text": question}]}],
+        inferenceConfig={"maxTokens": 64, "temperature": 0},
+    )
+    message = response["output"]["message"]
+    return "".join(block.get("text", "") for block in message.get("content", []))
+
+
+def main() -> None:
+    answer = answer_question("What is the capital of Australia? Reply in one sentence.")
+    print(answer)
+    braintrust.flush()
+
+
+if __name__ == "__main__":
+    main()

examples/bedrock_runtime/pyproject.toml

@@ -0,0 +1,12 @@
+[project]
+name = "braintrust-bedrock-example"
+version = "0.1.0"
+description = "boto3 Bedrock Runtime client traced with Braintrust"
+requires-python = ">=3.10"
+dependencies = [
+    "braintrust",
+    "boto3",
+]
+
+[tool.uv.sources]
+braintrust = { path = "../../py", editable = true }

py/noxfile.py

@@ -230,6 +230,19 @@ def test_cohere(session, version):
     _run_tests(session, f"{INTEGRATION_DIR}/cohere/test_cohere.py", version=version)
 
 
+BOTO3_VERSIONS = _get_matrix_versions("boto3")
+
+
+@nox.session()
+@nox.parametrize("version", BOTO3_VERSIONS, ids=BOTO3_VERSIONS)
+def test_bedrock_runtime(session, version):
+    """Test the boto3 Bedrock Runtime integration."""
+    _install_test_deps(session)
+    _install_matrix_dep(session, "boto3", version)
+    _install_matrix_dep(session, "botocore", version)
+    _run_tests(session, f"{INTEGRATION_DIR}/bedrock_runtime/test_bedrock_runtime.py", version=version)
+
+
 INSTRUCTOR_VERSIONS = _get_matrix_versions("instructor")
 
 

py/pyproject.toml

@@ -438,6 +438,14 @@ latest = "pytest==9.0.3"
 [tool.braintrust.matrix.braintrust-core]
 latest = "braintrust-core==0.0.59"
 
+[tool.braintrust.matrix.boto3]
+latest = "boto3==1.43.29"
+"1.34.116" = "boto3==1.34.116"
+
+[tool.braintrust.matrix.botocore]
+latest = "botocore==1.43.29"
+"1.34.116" = "botocore==1.34.116"
+
 # ---------------------------------------------------------------------------
 # Vendor packages — optional third-party packages the SDK can work without.
 # Keys are matrix keys; values are Python import names.  The noxfile uses this
@@ -461,6 +469,7 @@ agentscope = ["agentscope"]
 agno = ["agno"]
 autogen = ["autogen-agentchat"]
 anthropic = ["anthropic"]
+bedrock_runtime = ["boto3", "botocore"]
 cohere = ["cohere"]
 claude_agent_sdk = ["claude-agent-sdk"]
 crewai = ["crewai"]
@@ -488,6 +497,8 @@ anthropic = "anthropic"
 cohere = "cohere"
 autoevals = "autoevals"
 braintrust-core = "braintrust_core"
+boto3 = "boto3"
+botocore = "botocore"
 crewai = "crewai"
 dspy = "dspy"
 google-adk = "google.adk"

py/src/braintrust/auto.py

@@ -13,6 +13,7 @@
     AgnoIntegration,
     AnthropicIntegration,
     AutoGenIntegration,
+    BedrockRuntimeIntegration,
     ClaudeAgentSDKIntegration,
     CohereIntegration,
     CrewAIIntegration,
@@ -72,6 +73,7 @@ def auto_instrument(
     openai_agents: bool = True,
     cohere: bool = True,
     autogen: bool = True,
+    bedrock: bool = True,
     crewai: bool = True,
     strands: bool = True,
     temporal: bool = True,
@@ -106,6 +108,7 @@ def auto_instrument(
         openai_agents: Enable OpenAI Agents SDK instrumentation (default: True)
         cohere: Enable Cohere instrumentation (default: True)
         autogen: Enable AutoGen instrumentation (default: True)
+        bedrock: Enable boto3 Bedrock Runtime instrumentation (default: True)
         crewai: Enable CrewAI instrumentation (default: True)
         strands: Enable Strands Agents instrumentation (default: True)
         temporal: Enable Temporal instrumentation (default: True)
@@ -195,6 +198,8 @@ def auto_instrument(
         results["cohere"] = _instrument_integration(CohereIntegration)
     if autogen:
         results["autogen"] = _instrument_integration(AutoGenIntegration)
+    if bedrock:
+        results["bedrock_runtime"] = _instrument_integration(BedrockRuntimeIntegration)
     if crewai:
         results["crewai"] = _instrument_integration(CrewAIIntegration)
     if strands:

py/src/braintrust/conftest.py

@@ -247,6 +247,12 @@ def get_vcr_config():
             "openai-api-key",
             "x-goog-api-key",
             "x-bt-auth-token",
+            "x-amz-security-token",
+            "x-amz-date",
+            "x-amz-content-sha256",
+            "amz-sdk-invocation-id",
+            "amz-sdk-request",
+            "x-amzn-bedrock-api-key",
         ],
     }
 

py/src/braintrust/integrations/__init__.py

@@ -3,6 +3,7 @@
 from .agno import AgnoIntegration
 from .anthropic import AnthropicIntegration
 from .autogen import AutoGenIntegration
+from .bedrock_runtime import BedrockRuntimeIntegration
 from .claude_agent_sdk import ClaudeAgentSDKIntegration
 from .cohere import CohereIntegration
 from .crewai import CrewAIIntegration
@@ -29,6 +30,7 @@
     "AgnoIntegration",
     "AnthropicIntegration",
     "AutoGenIntegration",
+    "BedrockRuntimeIntegration",
     "ClaudeAgentSDKIntegration",
     "CohereIntegration",
     "CrewAIIntegration",

py/src/braintrust/integrations/auto_test_scripts/test_auto_bedrock_runtime.py

@@ -0,0 +1,43 @@
+"""Test auto_instrument for boto3 Bedrock Runtime."""
+
+import os
+
+
+os.environ.setdefault("AWS_EC2_METADATA_DISABLED", "true")
+if not os.environ.get("AWS_PROFILE") and not os.environ.get("AWS_BEARER_TOKEN_BEDROCK"):
+    os.environ.setdefault("AWS_ACCESS_KEY_ID", "testing")
+    os.environ.setdefault("AWS_SECRET_ACCESS_KEY", "testing")
+    os.environ.setdefault("AWS_SESSION_TOKEN", "testing")
+
+import boto3
+from braintrust.auto import auto_instrument
+from braintrust.integrations.test_utils import autoinstrument_test_context
+
+
+MODEL = os.getenv("BRAINTRUST_BEDROCK_CONVERSE_MODEL", "us.amazon.nova-lite-v1:0")
+REGION = os.getenv("AWS_REGION") or os.getenv("AWS_DEFAULT_REGION") or "us-east-1"
+
+results = auto_instrument()
+assert results.get("bedrock_runtime") is True
+
+results2 = auto_instrument()
+assert results2.get("bedrock_runtime") is True
+
+with autoinstrument_test_context("test_auto_bedrock_runtime", integration="bedrock_runtime") as memory_logger:
+    client = boto3.client("bedrock-runtime", region_name=REGION)
+    response = client.converse(
+        modelId=MODEL,
+        messages=[{"role": "user", "content": [{"text": "Say hello in one word."}]}],
+        inferenceConfig={"maxTokens": 20, "temperature": 0},
+    )
+    assert response["output"]["message"]["role"] == "assistant"
+
+    spans = memory_logger.pop()
+    assert len(spans) == 1, f"Expected 1 span, got {len(spans)}"
+    span = spans[0]
+    assert span["metadata"]["provider"] == "bedrock"
+    assert span["metadata"]["model"] == MODEL
+    assert span["metadata"]["endpoint"] == "converse"
+    assert span["span_attributes"]["name"] == "bedrock.converse"
+
+print("SUCCESS")