The well page pre-loads imagery for the observations and wells around the one you're viewing, so navigation feels instant. It was also downloading full-resolution images for every site in each of those observations even when a stitched montage was already available and the page never needed them — several megabytes per observation.

Those downloads now happen only when the stitched view isn't there and the page has to assemble the montage itself.

Retrying an image classification used to redo the entire plate, repeating work on wells that had already succeeded. Retries are now targeted: only the wells without a successful result for the selected model are processed again, so a retry after a partial failure is much quicker and doesn't discard good results.

Classification results are also more durable. Previously, an error arriving late in a run could overwrite results that had already come back successfully. Individual well outcomes are now retained even when a later step reports a problem, so a run that partly succeeded shows you exactly which wells worked and which didn't.

Teams working from sites geographically distant from their database region saw file uploads take much longer than expected, because each batch of upload status updates was committed separately and every commit meant another round trip across regions.

The agent now applies a run of file status updates together in a single transaction, cutting the number of round trips dramatically. In testing, a sync request covering 50 files went from roughly 11 seconds to under 2 seconds at the network latency typical of our Tokyo region. The improvement scales with distance, so the further a site is from its database, the more noticeable it is.