VivekChoudhary77’s JEV-AWS-Agent uses Jev to select actions on predefined AWS Console tasks. Playwright controls the browser; separate AWS API checks inspect the result. The repository’s README contains only the project name, so the implementation is the primary explanation.
Turn a page into choices
- The agent collects visible controls, labels and state, then asks Jev to choose an operation and, when needed, a target.
- For text entry, a separate choice maps the selected field to a predefined task input. Large control lists are scored and shortlisted first.
- Low-confidence decisions go to a human for selection or abort. A focused yes/no question can help the agent decide whether a task appears complete.
Verify beyond the browser
- The task file covers budget, Lambda, EC2 and Bedrock workflows; its RDS creation task is disabled. Some tasks create real, potentially billable resources.
- After a claimed completion, task-specific checks query AWS APIs for properties such as budget settings or a function URL’s authentication type. Resource-creating tasks also define cleanup for names associated with the run.
- The Bedrock check looks for a model invocation metric in a time window. It does not prove which prompt caused that invocation or independently check the reply.
Treat the gates as guardrails, not guarantees
The code restricts console URLs and regions, blocks selected sensitive pages, and asks for approval on controls whose names match a list of consequential actions. A dry run stops at an approval gate. A keyword gate can miss consequential actions with different labels; confidence and typed output do not establish a correct or safe decision.
Verification: inspected the README, task definitions, agent and verifier code. The project was not run, its tests were not reproduced, and this is not a security audit. No end-to-end success rate is established here.