Rovo Agent built in Studio fails to run in Automation.

Question

I have rovo chats that I am able to execute in Studio/Chat with the correct inputs.  However when I use the same input in an automation, the the Rovo chat fails to return data.  Is there documenation available on how to use Rovo studio agents successfully (self created) in automations?

Josh · Answer

One extra diagnostic is to compare the exact payload reaching the rule with the payload you send in Studio. Add a Log action immediately before Use agent and log the issue key plus each prompt variable (not secrets), then log {{agentResponse}} immediately after it. In Automation, smart values are evaluated in the rule context and missing fields often become empty strings, while Studio may be supplying a remembered input. Also verify the rule actor has product access and permission to the project, issue, and any Confluence/knowledge source; test the same URL as that actor. Once the response is stable, keep the agent contract small (for example {"category":"...","summary":"..."}), validate required keys before using them, and fail with a visible audit-log message rather than creating a partial update.

Sami Shaik · Answer

@Ken Young , 
@Gerardo Martín JSON contract and @Mihai Perdum actor point are both right, and they describe two different failures that look identical from the outside: an empty response. Here is how to tell them apart in one run, then the design that stops it happening again.

The five-minute test. Open the rule's audit log for a failed run and expand the Use agent action.

* If the action shows success with an empty or near-empty response, it is the actor problem Mihai described. The agent ran, but as the rule's actor it could not see the pages, work items or knowledge sources it reads for you in chat, so it answered from nothing. No error is raised because, from the agent's point of view, nothing went wrong.
* If the action shows the agent's full response text but the next action gets blank values, it is the smart-value problem (AUTO-1873 / AUTO-1897): copy {{agentResponse}} into a Create variable first, as Mihai said, and read the JSON fields from the variable.
* If the action itself errors, read the message; that is usually a permission or a licence on the actor, not the agent.

Then prove the actor theory rather than guessing. Note the account shown as Actor in Rule details. Log in as that account (or ask its owner to) and open, by hand, one page or work item the agent's answer depends on. If that account cannot open it, you have your cause; if it can, look at the agent's knowledge sources in Studio, because a source restricted to a space the actor is not in fails the same way.

The design that holds:

* Run agent rules under a dedicated automation actor, a named account with explicit, reviewable access to exactly the spaces and projects the agent needs, not a human admin's account. When that person leaves or changes teams, your rules keep running and nobody has to guess what the agent could see.
* Put a condition before Use agent that checks the input exists (the work item field or page the prompt relies on), so the agent is never asked to reason about nothing.
* Follow Gerardo's contract: the agent returns a fixed JSON shape, and the rule fails loudly with a Log action when a required field is missing, instead of silently writing blanks.

On documentation: the page Mihai linked is the official one, and it covers the action and the output formats. It does not cover the actor's visibility, which is why this thread exists; treat that as the part you have to design yourself.

Mihai_LeanZero · Answer

There's an official page for this. "Automate Rovo agents" on support.atlassian.com/rovo/docs/agents-in-automations/ covers the Use agent action, and the automation smart values page lists the output formats ({{agentResponse}}, .asString, .asObject etc).

First thing I'd check is who the agent runs as. In chat it's you. In a rule it's the connection user, Atlassian says on AUTO-2337 that's intended because "automations are designed to run when there is no user present", so if that account can't see the pages or work items the agent reads it probably comes back empty with no error.

Second one. If the agent answers fine but text functions on it like .split() or .length() come back blank in the next action, that's AUTO-1873 / AUTO-1897. The workaround on 1897 is copying {{agentResponse}} into a Create variable action first.

Which account is the rule's connection set to?