Sandbox: AWS Bedrock AgentCore

AWS Bedrock AgentCore provides a managed code interpreter backed by AWS infrastructure.

Prerequisites

• AWS credentials with access to bedrock-agentcore

• Bedrock AgentCore available in your region (default: us-east-1)

AWS credentials are read from:

1. Parameters passed directly to the constructor (aws_access_key_id, aws_secret_access_key)

2. Environment variables: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION

3. Standard boto3 credential chain (IAM roles, ~/.aws/credentials, etc.)

Setup

from gllm_tools.code_interpreter.code_sandbox.bedrock_agentcore_sandbox import BedrockAgentCoreSandbox

# Using environment credentials
sandbox = BedrockAgentCoreSandbox(region="us-east-1")

# Using explicit credentials
sandbox = BedrockAgentCoreSandbox(
    region="us-east-1",
    aws_access_key_id="AKIA...",
    aws_secret_access_key="secret...",
)

Note: The session is created lazily — no AWS connection is made until the first execute_code call.

Constructor Parameters

Parameter	Type	Default	Description
region	str	"us-east-1"	AWS region for Bedrock service
aws_access_key_id	str | None	None	AWS access key ID. Falls back to env vars / IAM role.
aws_secret_access_key	str | None	None	AWS secret access key. Falls back to env vars / IAM role.
language	str	"python"	Programming language (only Python supported)
session_timeout	int	60	Server-side session lifetime in seconds
aioboto3_session_kwargs	dict | None	None	Extra kwargs passed to aioboto3.Session (e.g., profile_name)

Timeout Behaviour

There are two independent timeouts:

Timeout	Parameter	Scope	Description
Per-execution	timeout in execute_code()	Client-side	Max time for a single code run (default: 30s)
Session lifetime	session_timeout in constructor	Server-side	Max total lifetime of the session (default: 60s)

Full Example

import asyncio
from gllm_tools.code_interpreter.code_sandbox.bedrock_agentcore_sandbox import BedrockAgentCoreSandbox

async def main():
    sandbox = BedrockAgentCoreSandbox(
        region="us-east-1",
        session_timeout=120,
    )

    try:
        result = await sandbox.execute_code("""
import math
print(f"Pi = {math.pi:.6f}")
print(f"e  = {math.e:.6f}")
""", timeout=30)

        if result.status.value == "success":
            print(result.stdout)
        else:
            print(f"Error: {result.error}")

    finally:
        await sandbox.terminate()

asyncio.run(main())

With SandboxCodeInterpreter

import asyncio
from gllm_tools.code_interpreter.code_sandbox.bedrock_agentcore_sandbox import BedrockAgentCoreSandbox
from gllm_tools.code_interpreter.code_interpreter.sandbox_code_interpreter import SandboxCodeInterpreter

async def main():
    sandbox = BedrockAgentCoreSandbox(region="us-east-1", session_timeout=300)
    interpreter = SandboxCodeInterpreter(sandbox=sandbox, lm_invoker=lm_invoker)

    result = await interpreter.execute("Compute the first 10 Fibonacci numbers")
    print(result.text)

    await sandbox.terminate()

asyncio.run(main())

Notes

• State does not necessarily persist between separate execute_code calls (depends on AWS session state).

• File upload via the files= parameter is supported.

• download_file(path) returns bytes for both regular files and images.

• Using an AWS profile: pass aioboto3_session_kwargs={"profile_name": "my-profile"}.