Async PDF queue refactors are exactly the kind of backend plumbing AI handles well.

This is a standard architectural pattern — offload slow work to a task queue, poll for status, store output in object storage. The structure is nearly identical across FastAPI projects and well-represented in training data.

Success criteria are concrete and testable: requests no longer timeout, a polling endpoint returns job status, and completed PDFs are accessible via signed S3 URLs. There is minor ambiguity around Celery vs. RQ choice and error-state handling, but these are resolvable decisions.

The agent needs access to the existing codebase, Jinja2 templates, Weasyprint setup, S3 bucket credentials, and the current FastAPI app structure. Without these, the agent produces generic scaffolding that may not integrate cleanly — the user must supply this context.

Bugs in the implementation could break PDF generation entirely or expose S3 URLs incorrectly, but this is a dev/staging change that can be tested before production deployment. No irreversible data loss or security breach is likely if reviewed before merge.

Choosing between Celery and RQ, setting TTLs on signed URLs, and structuring the polling response schema are minor decisions an agent can make with reasonable defaults. No taste, ethics, or relationship context is needed.