AWS Bedrock

What this integration actually does

AiraBedrockHandler wraps a boto3 Bedrock client so every call to invoke_model and invoke_agent runs through aira.authorize() first. If the policy engine denies, the wrapped method raises AiraInvocationDenied and Bedrock is never called — no tokens billed, no side effects.

Successful calls are notarized with outcome="completed" and failed calls (any exception from the underlying Bedrock call) are notarized with outcome="failed" so the action transitions correctly in Aira.

Install

pip install aira-sdk[bedrock]

This pulls in boto3 as a peer dependency. If you already have boto3 installed you're covered.

Full example — gated invoke_model

import boto3
from aira import Aira
from aira.extras.bedrock import AiraBedrockHandler, AiraInvocationDenied

aira = Aira(api_key="aira_live_...")
handler = AiraBedrockHandler(client=aira, agent_id="support-agent")

# Create the raw Bedrock client
bedrock = boto3.client("bedrock-runtime", region_name="eu-west-1")

# Wrap it — this is where authorize() runs
bedrock.invoke_model = handler.wrap_invoke_model(bedrock)

try:
    response = bedrock.invoke_model(
        modelId="anthropic.claude-3-5-sonnet-20241022-v2:0",
        body='{"anthropic_version": "bedrock-2023-05-31", "max_tokens": 1024, "messages": [{"role": "user", "content": "Draft a refund reply"}]}',
    )
    print(response["body"].read())
except AiraInvocationDenied as e:
    print(f"Aira blocked: {e.code} — {e.message}")

Full example — gated invoke_agent

import boto3
from aira import Aira
from aira.extras.bedrock import AiraBedrockHandler, AiraInvocationDenied

aira = Aira(api_key="aira_live_...")
handler = AiraBedrockHandler(client=aira, agent_id="support-agent")

bedrock_agent = boto3.client("bedrock-agent-runtime", region_name="eu-west-1")
bedrock_agent.invoke_agent = handler.wrap_invoke_agent(bedrock_agent)

response = bedrock_agent.invoke_agent(
    agentId="AGENT123",
    agentAliasId="TSTALIASID",
    sessionId="sess-001",
    inputText="Process refund request RMA-9876",
)

What happens when a policy denies

1. Your code calls bedrock.invoke_model(modelId=..., body=...).
2. The wrapper calls aira.authorize(action_type="model_invoked", details="Bedrock invoke_model: anthropic...").
3. Policy engine runs. If denied, the wrapper raises AiraInvocationDenied.
4. Bedrock is never called. No tokens billed, no response generated.
5. The denied PolicyEvaluation row persists for audit — you can see exactly why this invocation was blocked.

Action types

The wrapper uses different action_type strings so you can write policies targeting one or the other:

Write a policy like:

{
  "mode": "rules",
  "conditions": [
    {"field": "action_type", "op": "eq", "value": "model_invoked"},
    {"field": "details", "op": "contains", "value": "anthropic.claude-3-opus"}
  ],
  "decision": "require_approval"
}

...to require human approval for any call to a specific model family without affecting other Bedrock calls.

When to use this vs inline aira.authorize()

Known limits

• The wrapper replaces the method in place. After bedrock.invoke_model = handler.wrap_invoke_model(bedrock), any subsequent code that calls bedrock.invoke_model will go through Aira. If you have existing code that relies on the unwrapped method, use a different boto3 client instance.
• Details strings do not include the prompt body. By default we use f"Bedrock invoke_model: {modelId}" to keep auth latency low and avoid sending the full prompt to Aira's details column. If you want to gate on prompt content, pass a details builder callback — open an issue if you need this, it's a small addition.
• Streaming responses are supported because we only wrap the call site, not the returned stream. You get one authorize + one notarize per stream open.

Proof it works

• Python: tests/test_extras_bedrock.py (132 lines)
• Pinned in the INTEGRATIONS registry (aira/extras/__init__.py) as kind="gate", pre_execution_gate=True.

Related

• Policies — configuring model-specific policies
• Content scan policies — regex-based scan over action details