Agentic Workflow returns duplicate responses intermittently
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
yesterday
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
45m ago
Duplicate replies usually happen because more than one part of the workflow is speaking to the user. There's no single switch for "one final response," but you can control who speaks and when. Field names vary by release, so confirm them in your instance.
Start by finding where each duplicate comes from.
- Go to AI Agent Studio and open your agentic workflow.
- Run it from the Testing tab until a duplicate appears.
- Open the execution plan and check which step produced each copy.
The most common cause is a tool sending output to the user while the agent also summarizes it.
- Open each agent's tools and check the output setting.
- Set tools to return output to the agent only.
- Let the agent write the single final reply.
If your workflow has several agents, each one may be sending its own closing message.
- Pick one agent to give the final reply.
- In the other agents' instructions, tell them not to address the user.
Intermittent duplicates often mean the agent retried a step because a tool was slow, failed, or returned unclear output.
- Check the execution plan for the same tool running twice.
- Test that tool on its own and fix any errors or timeouts.
- Make the tool return a short, clear success message.
- Add "Give one final answer and do not repeat earlier messages" to the agent's instructions.
If duplicates show up only in your live chat, a Virtual Agent topic may be answering alongside the agent.
- Compare results in the studio test panel and your live channel.
- Check Virtual Agent topics for overlapping triggers.
- Search Now Support for known errors in your release.
Test these changes in sub-production first. If it still happens, share your release, the number of agents and tools, and the channel you're using.
