Runs

Each execution of a Workflow produces a run with its own record: the input received, the steps taken, what each node returned and the final result.

This is the practical difference between a Workflow and an Agent when you investigate. An Agent gives an answer; a Workflow leaves a trail of how it got there.

What a run records

• Input: the values the person filled in on the form.
• Intermediate steps: what each node received and produced, in execution order.
• Output: the result shown to the person who ran it.
• Errors: which node the run stopped at, when it stopped.

Investigate a wrong result

The trail makes investigation mechanical. Walk the nodes backwards until you find the first one whose output is already wrong. The problem is in that node or in the input it received.

The two most common cases:

• The vector node returned nothing: the Collections configured in it are empty, or the documents never reached Active.
• The LLM node returned something outside the expected format: the node's prompt isn't explicit about the format, and the next node can't consume the output.

When a run fails

A failure stops the flow at the problematic node and records the error there. The following nodes don't run.

Failures in integration nodes are usually external: an expired credential, an endpoint that is down, a domain that isn't allowed. Failures in logic or code nodes are usually configuration.

Version that ran

A run keeps the version of the Workflow that was published at the time it ran.

This matters when you investigate an old case: the flow may have changed since then, and what you see on the canvas today isn't necessarily what ran.

Next steps

• Adjust the flow in Build on the canvas.
• See the model consumption of these runs in Observability.