Troubleshoot agents

This page covers common issues you may encounter when building or running agents, along with steps to resolve them.

Agent errors

Generic "Something went wrong" or request ID errors

If your agent displays a generic error with a request ID, try the following:

1. Open a new chat session and run the agent again. Some errors are temporary.
2. If the error persists, check whether your agent's LLM model is correctly configured in the Admin Console under Platform → LLM.
3. Downvote the response and include a brief description of the issue. This helps your admin and Glean Support in their investigation.

Agent times out or doesn't complete

Agents with many steps or large amounts of context may time out. To reduce the chance of timeouts:

• Reduce the number of steps. Combine related logic into fewer steps where possible.
• Use a Plan and Execute step. Replace chains of Company Search, Think, and Respond steps.
• Limit context per step. Set memory to "No memory" on steps that perform unrelated tasks, and manually tag only the steps whose output is needed.
• Constrain output length. In Respond step instructions, specify a maximum length for the response.
• Try a faster model. If latency is contributing to timeouts, experiment with a faster model while monitoring output quality.

Scheduled run fails but manual run succeeds

If an agent completes successfully when you run it manually but stops part-way through when it runs on a schedule, the scheduled run is most likely hitting the scheduled run execution limit. Manual runs are not subject to this limit, which is why the same agent can finish manually and fail on a schedule.

To resolve this:

• Review the agent for steps that can be combined, batched, or moved to a faster model. See Scheduled run execution limit for detailed strategies.
• Split long reports into multiple scheduled agents that each finish within the limit.
• Confirm that every write tool in the agent is enable tools that still require per-run confirmation.

Chat session stops responding after a long conversation

If a chat with an agent worked initially but later turns consistently fail with a generic error in the same session, the conversation has likely exceeded the chat session size limit.

Start a new chat session to recover. If you need to keep context, ask the agent in the original session to summarize the prior conversation, then paste that summary into the new session as your first message.

Inconsistent agent outputs

If your agent produces different results across runs:

• Review the agent's steps in debug mode to identify which step produces inconsistent output.
• Make sure that Company Search steps are scoped narrowly with specific filters rather than broad queries.
• Avoid using "As many as possible" for the number of search results, as this can overload context and lead to truncation.
• Verify that all referenced documents are indexed and accessible to the user running the agent.

Company Search limitations

• Company Search returns a maximum of 500 results per step. If your use case requires processing more documents, break the work into multiple scoped searches.
• Using "As many as possible" combined with "Entire document content" can cause context overload. Use specific queries and limit the number of results to improve reliability.

SharePoint content in agents

Glean indexes individual documents stored in SharePoint document libraries but doesn't index SharePoint view pages (URLs ending in AllItems.aspx or DocSetHome.aspx). These pages are UI containers, not content items.

If your agent uses a Read document step pointed at a SharePoint view URL, it returns an "unable to read" error. To work around this:

1. Replace the Read document step with a Company Search step.
2. Scope the Company Search to the specific SharePoint library or folder.
3. The agent retrieves and operates on the individual documents within that library.

For more details on what SharePoint content Glean indexes, see SharePoint known limitations.

Citations not appearing

Citations appear when the response relies on enterprise sources. If you notice citations are missing:

• Rephrase your question to reference company-specific knowledge. For example, add context like "According to our FY25 strategy doc..."
• If citations stopped appearing after enabling the new agentic engine, ask your admin to verify that the agentic reasoning model is set to a supported model in Admin Console > Platform > LLM.
• For API integrations, make sure requests use agentConfig.agent = "ADVANCED" rather than "DEFAULT". The DEFAULT config uses the legacy assistant, which may not return citations consistently.

For more information about how citations work, see Citations.

See also

• Debug your agent
• Create a more powerful agent
• Create agents with Auto mode