Decide which access requests an agent may approve

Your security team drowns in access requests, and most of them are routine. An agent may approve one narrow, pre-approved class of request on its own. Everything else goes to a security engineer, even the requests it would deny.

Write triage_access_requests(scopes, environments, durations_days, peer_scopes, max_auto_days): the rule your agent follows for each request.

• scopes (list of str): the permission each request asks for, such as "read-logs".
• environments (list of str): the environment it applies to.
• durations_days (list of int): how many days the access should last.
• peer_scopes (list of list of str): for each request, the scopes the requester's peers already hold.
• max_auto_days (int): the longest duration the agent may approve.
• Index i in each list is one request. Return one decision per request, in order: "auto_approve" or "human_review".

Rules, for each request. Names match exactly, so case matters, and the first rule that matches decides:

1. The environment is not dev, staging or sandbox: "human_review".
2. The scope is not read-logs, read-metrics or sandbox-deploy: "human_review".
3. The scope is not in that request's peer_scopes: "human_review".
4. The duration is below 1 or above max_auto_days: "human_review".
5. Otherwise: "auto_approve".

Then, in a sentence or two: why does the agent send even clear denials to a person instead of denying them itself?