Training
The training dashboard lives at Browser Agents > [Agent] > Training. It has four tabs: Datasets, Runs, Metrics, and Analytics.
Datasets
A dataset is a collection of training examples used to improve agent behavior. Each dataset has a status:
| Status | Meaning |
|---|---|
| Draft | Being built, not yet used for training |
| Ready | Validated and available for training jobs |
| Archived | No longer active |
From this tab you can create, edit, and delete datasets. A dataset's examples come from labeled agent runs (see Labeling below).
Training Runs
The Runs tab shows the history of training jobs for this agent.
To start a new training run:
- Select a dataset with Ready status
- Choose configuration options
- Start the run
Each run shows a progress bar and links to training logs. Training jobs improve the agent's decision-making based on the examples in the selected dataset.
Metrics
The Metrics tab shows agent performance over time. Use the time period selector (7d / 30d / 90d) to adjust the window.
Tracked metrics:
- Success rate
- Error rate
- Average steps to success
- Coverage score
Charts show trends across runs so you can see whether training is improving performance.
Analytics
The Analytics tab provides a comparison view: improvement delta vs the previous snapshot, and evaluation set performance broken down by scenario type.
Golden Dataset
Each agent has a Golden Dataset — a curated set of high-quality training examples extracted from labeled runs. This is the source of truth for training. Build the golden dataset by labeling runs and promoting good examples into it.
Labeling
Labeling is in development. The concept: review agent run steps, mark each step as correct or incorrect, and add corrections where needed. The output feeds into training datasets. This UI is not yet available.
Back to Overview