Run a Batch

One successful chat does not show how an agent behaves across varied inputs. A batch runs every row in a dataset through the same Flow and applies your evaluations consistently.

Prepare representative data

Create a Dataset with realistic user inputs. Include expected answers or labels when an evaluation requires a reference value. Start small and cover normal requests, edge cases, and questions the agent should decline or say it cannot answer.

Create the batch

1. Open the Flow's Batch tab.
2. Select Create Batch.
3. In Data and Evals, choose the Flow, dataset, and evaluations.
4. In Map Inputs, map dataset columns to the Flow and evaluator inputs.
5. In Review and Run, verify the configuration and start the batch.

⤢Expand

Screenshot previewThe Create Batch dialog on the Data and Evals step with Flow, dataset, and evaluation selectors

×

Use the results

Compare responses and evaluation scores across rows. Open the trace for a failed or surprising row to determine whether the issue came from retrieval, instructions, tool use, or the model response.

After changing a prompt, model, memory setting, or tool, rerun the same dataset. Keeping the inputs and evaluations stable makes the effect of each change visible.

See Batches for result analysis and run comparison.