Why is my Rovo Agent saying that? Let's Debug Smarter.

[#1]

You built your Rovo Agent, tested a few prompts, and something feels… off. Maybe it's pulling the wrong information. Maybe it's ignoring instructions you clearly set. Or maybe it's triggering the wrong subagent entirely. Whatever the symptom, the frustration is the same — "Why is my agent saying that?". 
Let's see how we can find an answer to that question.

🕵🏻‍♂️ What to look for when debugging

• Missing or wrong information? If an answer lacks context, check the knowledge section in studio to see what your agent referenced. If the sources are wrong, check that the agent has the correct knowledge sources and tools connected.

• Conflicting Agent instructions? Sometimes less is more! If agent instructions are complex, check if any of these are conflicting with each other.

• Wrong subagent triggered? If the agent used a different subagent than you expected, your instructions or trigger phrases may be too broad or overlapping. Tighten the wording.

💬 Chat with Rovo agent to debug itself

Often the quickest fix is asking the agent itself questions. A few habits that consistently get more meaningful answers:

• Be specific. Mention your hardline criteria or specific detail that is missing/incorrect.

• Give context. Include part of the response that you are not happy about.

• Ask one thing at a time. Break multi-part requests into steps.

• Ask it to explain itself. This surfaces gaps in what the agent can see.

Example scenarios & prompts to try:

Scenario	Example prompt to the agent
The agent picked the wrong approach	"Which subagent or skill did you use for this, and why did you choose it?"
The answer looks wrong or may be only partially correct. You want to know why	"What information and sources did you use to answer that? List them."
The response feels incomplete	"What assumptions did you make, and what would you need from me to give a more complete answer?"
The agent sent correct response but the response format is unexpected.	“Which agent instructions led you to generate response in so&so format”? or ” Which agent instructions stopped you from generating response in JSON format”?

 📺 Use the built-in Debug view

Rovo Chat has a built-in Debug responses from Rovo agents feature that lets you peek "under the hood" of any agent reply.

What you'll see:

• Subagents used — which skill/instruction set the agent used to answer.

• Agent ID and Request ID — unique references for that response.

• Skills used - List of skills used by the agent to provide a response.

Tip: The Debug button only appears on a message from a custom agent — not in a regular Rovo chat. If you don't see it, confirm you're chatting with a custom Rovo agent.

🔬 Click here to Deep Dive into Debug Response!

Expanding Subagents used tab:

Check which subagents the response used. The Input and Output tabs show the subagent query and its actual response. If the correct subagent triggered but the agent’s final output differs, this can explain why.

Compare the subagent output with the agent response to see the differences.

Expanding Skills used tab:

This tab lists the tools Rovo uses to generate responses, for example Get work item (Tool ID = jira-get-isssue-full) and Content Read (Tool ID = ContenReadTool). Users can pinpoint the exact skills used or missing, provide micro input for each skill, and view the corresponding outputs.

This view helps users determine whether the agent is unable to perform a task, or missing a specific micro detail.

🪜 Example Debug Flow

🎫 Last Resort : Raise a new support ticket

If a response still looks wrong after checking the above, open a support ticket and use Copy all button to send Agent, Agent ID, and Request ID from the Debug view. These details let Atlassian trace the exact response quickly.

🔗 Atlassian support portal: support.atlassian.com/rovo

[Alex Gallien]

Great content @Sumit Fartode - thanks for taking the time to write this up!

[Esther Ortega]

Great article, Sumit!

I really like the diagnostic checklist: checking that connected knowledge sources are correct, spotting conflicting instructions within the agent, and verifying whether the wrong subagent was triggered due to overlapping trigger phrases. It nails exactly the points where most time gets lost.

But what I like most is asking the agent directly "which subagent did you use and why?" or "what sources did you reference?": a simple meta-prompting technique that anyone can apply right away, with no extra tools, and that helps the agent itself surface its own gaps in reasoning or data access. Combined with Rovo Chat's debug view (subagents, skills/tools used, Agent/Request IDs), it turns what used to be a black box into something far more traceable.

Looking forward to a follow-up with complex real-world examples. Thanks for sharing!

[MeghnaP_LogicLemur Labs]

@Sumit Fartode  One thing I’d add to this excellent debugging flow is historical visibility.

The Debug view is great when you’re investigating one response: which subagent ran, which skills were used, what the request ID was, etc.

But once agents are used regularly, another question becomes important:

“Is this a one-off failure, or is my agent consistently making the same mistake?”

For example:

• Is the same subagent being triggered incorrectly multiple times?

• Are certain skills failing repeatedly?

• Are responses becoming less accurate after an instruction change?

• Which agent executions are taking longer or producing unexpected outcomes?

That moves agent debugging from “debug this response” to “understand agent behavior over time.” when they assigned to any Jira issues.

I think that historical execution view will become increasingly important as teams move from experimenting with Rovo Agents to relying on them in day-to-day workflows.