For AI agents: a documentation index is available at the root level at /llms.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
Export to cloud storage. The Export to cloud storage node writes a dataset’s images, and their annotations in Supervisely, COCO or YOLO format, to an S3-compatible bucket. If files exist chooses whether to overwrite, skip or stop on files already in the bucket.
The agent runs flows. An AI agent can now start a flow run, follow its status, wait for it to finish, and read the task logs, reports and report files it produced. Before starting, it names any upstream node that has not produced a value yet.
Hardest images. The Evaluate object detection report shows the images with the lowest F1 first, each opening the ground truth beside the predictions. Sample images similar to failures starts its search from them.
Improvements
Labeling, Review and Consensus review start with default Outcome statuses: Submit and Skip for Labeling, Accept and Reject for the review nodes. A status left without a dataset gets a new one on the node’s first run.
A dataset input wired from another node’s output uses the dataset as it is when the node runs, instead of the version the producing node wrote. The node’s preview and sidebar show the dataset it will run on. The dataset picker can now select the first version.
Align layout on the canvas uses a new layered layout, and parallel connections no longer draw over each other or over nodes. Nodes the agent adds are placed the same way.
The agent’s flow edits move only the part of the flow they change, and it can add notes and change connections.
Prelabel with live training previews its output dataset on the node card.
Class colors of YOLO annotations no longer change between views.
Publish… in the flow menu is disabled outside the public flows workspace and says why, instead of opening a dialog that fails.
On OvalBee Cloud, a new personal workspace starts on the Pro plan during the beta.
Fixes
Train/validation split applies its default ratio and shuffling again. A node added while they were missing keeps the empty values, so add it again.
Task history opens the datasets a past run produced, not their newest version.
The plan shown after signing in as another user in the same tab is no longer the previous user’s.
Removed
The Dataset source verification node is gone, and Dataset preview no longer appears in the node picker. The Dataset node shows a dataset’s assets and annotations on its card.
