AWS has published a reference architecture for running Claude Code as a remote coding agent on Amazon Bedrock AgentCore. A developer can hand off a scoped task, close the laptop and later receive a GitHub pull request containing the proposed change and its verification result.
Each task runs in a separate microVM inside the customer’s AWS account. The agent clones an approved repository with a short-lived GitHub App token, plans with one model, writes and tests with another, and performs a review before opening the pull request. It can repeat the plan-and-review loop up to three times. A failed verification leaves the pull request as a draft, and the agent never merges.
Prompts are screened before entering the queue, model calls use private Bedrock endpoints, and long-lived GitHub credentials are not placed in the container. The sample can remember repository-specific lessons from earlier tasks while isolating each task’s filesystem and runtime.
AWS says the pattern suits bounded features, bug fixes and tests in trusted repositories. It is not intended for untrusted code or open-ended work, because repository commands still execute inside the agent environment and costs can be difficult to predict.