From the source
Amazon's guide shows how to run Claude Code as a remote agent on Amazon Bedrock AgentCore, so a coding task runs in its own isolated session inside the customer's AWS account.
- The sample, sample-remote-coding-agent, runs each task in its own AgentCore runtime session and returns a
GitHub pull request for review and merge.
- Teams pick Amazon Bedrock as Claude Code's model provider for billing through an existing AWS account with no per-seat licenses, audit logging in AWS CloudTrail, and geographic inference profiles for data residency.
- Amazon Bedrock doesn't use content to train models or share it with model providers, and providers can't access prompts, completions, or service logs, according to the post.
- Claude Code calls Amazon Bedrock with temporary IAM credentials, and AWS PrivateLink gives a VPC a private connection to Amazon Bedrock.
- The post says Mantle, the inference engine for Amazon Bedrock, is built for zero operator access, so AWS operators have no technical means to reach customer data, and zero data retention can be configured.
- The sample suits scoped features, bug fixes, and tests on a trusted repository, but not untrusted code, because the agent runs the repository's own commands, and not open-ended work, because each task stops after three attempts.






