AWS Details Remote Coding Agent Architecture Running Claude Code on Amazon Bedrock AgentCore
AI assistants · AI coding · Enterprise AI
AWS published an architecture sample showing how to run Claude Code as a remote coding agent on Amazon Bedrock AgentCore within private AWS environments.
AWS outlined an architecture and published an open-source sample demonstrating how to execute Claude Code inside Amazon Bedrock AgentCore runtime sessions within a user's AWS account. The implementation isolates each development task inside a dedicated microVM with separate CPU, memory, and filesystem. The agent coordinates planning with Claude Opus and code implementation using Claude Sonnet, runs repository verification commands, and outputs results as draft or standard GitHub pull requests without automatic merging.
Key takeaways
- Tasks run inside isolated AgentCore runtime microVM sessions within a customer's VPC, accessing Bedrock and AWS services via private VPC endpoints.
- The agent operates with short-lived GitHub user-to-server tokens managed via AgentCore Identity, restricting repository permissions without storing long-lived credentials.
- Bedrock Guardrails screens incoming task prompts for jailbreak attempts, prompt injection, and secrets before queuing tasks in DynamoDB and SQS.
- Tasks autonomously run repository setup and test suites up to three times, recording build lessons into AgentCore memory, before opening draft or standard pull requests.
Source：AWS Machine Learning Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.