the local chromium ceiling section is exactly what we hit last year. started with 3-4 parallel workers on a self-hosted runner, everything fine. scaled to 15 and tests started timing out randomly. spent two weeks debugging "flaky selectors" before realizing the runner was just OOMing silently.
the one-line swap from launch to connectOverCDP is the part i wish someone had told me earlier. in my head "moving to cloud browsers" sounded like a weekend migration project. ended up being a 30 minute change.
one question though - for long-running agent loops, how are folks handling session cleanup when the agent crashes mid-run? curious if providers auto-terminate or if you eat the idle time until timeout.
Priya Sharma
Backend dev obsessed with distributed systems
Good one. the stagehand lock-in point is underrated. teams pick it because the AI layer looks magical in a demo, then six months later they're stuck because half their agent logic is written against a proprietary API. plain CDP ages better.