Playwright Scripts Beat Screenshot-Heavy PDF Agents

The local-vs-cloud model choice is mostly a red herring; the screenshot spam comes from the agent layer, not from the model itself. For this workload, a scripted browser workflow with saved auth state is the right center of gravity.

• –Playwright can persist cookies and localStorage, so the two-page login problem is usually solved by logging in once, saving storage state, and reusing it in later runs.
• –A 1,000-PDF job is better treated as a batch script than an agent task: recurse through subsections, wait for selectors, trigger downloads, and checkpoint progress so failures resume cleanly.
• –If the site exposes direct PDF URLs or download endpoints, extracting those is far more robust than simulating human clicks through nested sections.
• –A hybrid RPA-plus-script setup only makes sense if RPA is confined to one-time MFA or login handoff; otherwise it adds another brittle layer.
• –Playwright’s download primitives make the download side straightforward once the authenticated session is stable, which is the real bottleneck here.