Agent output contracts and branch-specific outputs
An agent's outputs define what a run must deliver when it finishes: a text answer, a file, or a table. An agent has either one output for every run, or one output per Branch path so the deliverable matches the path the run took. Before a run can finish cleanly, we check that what it produced matches the output for that run.
Output formats
Each output has a Format:
Format | What the agent must deliver |
|---|---|
Text | A written answer in the run, with no generated or uploaded file. |
File | A document saved to August, such as a Word memo or an Excel workbook. |
Table | A structured table of results. |
One output or one per path
When every run produces the same deliverable, give the agent a single output. When different Branch paths produce different deliverables, add an output for each path instead, for example a File on the path that drafts a memo and Text on the path that only reports "no issues found". Only the output on the path the run took is checked; outputs on skipped paths are ignored.
Set up outputs
Outputs are edited on the canvas in edit mode. You can also ask for them in chat when you create or edit the agent.
On the agent page, open the menu next to Edit and choose Edit manually.
Select an output node, or click Add output to add another.
Fill in Output name, Format, and Description.
In Comes from step, choose the step that produces this output. For a per-path output, choose a step on that path.
Click Save, then test the agent with inputs that exercise each path.
The canvas editor is covered in Build and edit agents on the canvas.
How outputs are checked
When the agent tries to finish a run, we compare what it produced with the run's output. If they don't match, the agent is told what is wrong and tries again, rather than the run finishing with the wrong kind of deliverable. Typical mismatches:
A file was produced for a Text output.
A File output was expected, but no file was saved.
The run produced no result at all.
A run that finishes Completed with warnings skips this check. The warning explains what the agent could not deliver, so review it before relying on the result.
What to read next
Agents Overview for building and running agents
Agent Control Steps for Branch steps and their paths
Resume an agent run waiting on human input for runs paused on a Question step