Annotation task assignment

When new data lands in a V7 Darwin dataset, a Beam agent reads the batch metadata and the project's labeling stage. It applies your routing rule to distribute items to the right annotation workflow and updates task status and assignee fields where the connector allows it. Standard batches move into the queue without manual sorting. When a batch is missing labels, has an unexpected class, or exceeds a size threshold you set, the agent holds it and notifies a project lead. A person confirms how those items should be handled before any labeling actually begins.

Label quality checks

On completion of a labeling stage, a Beam agent reads the annotations and the review status inside V7 Darwin. It applies your acceptance rule, comparing labels against the criteria you set, then advances passing items and writes a review note back to each task. Clean work progresses to the next stage automatically. Items that fail a check, show disagreement between annotators, or fall below your confidence threshold are flagged for a human reviewer instead of auto-approval. Your team keeps the final say on whether contested labels enter the training set, and the agent records each decision.

Dataset export readiness

Before a training run, a Beam agent reads a V7 Darwin dataset's task states and completion counts. It applies your readiness rule to decide whether the set is fully labeled and reviewed, then updates a status field and notifies the ML team when the export criteria are met. Datasets that clearly meet the bar are marked ready. When required classes are underrepresented, tasks remain open, or counts look off against your threshold, the agent routes the dataset to a person rather than signaling go. Someone verifies the gaps before the data feeds a model.