Skip to Content
Getting StartedCommon tasksWhy is my agent failing?

Why is my agent failing?

Start at the Dashboard, find the agent whose success rate has dropped, then open the failed runs that share the same error and read the step where they stop. Morse groups failures for you and ranks the groups by what they cost, so the first group is the one to read first. What you get is a pattern to check in the trace, not a proven cause.

Steps

  1. Find the agent. Open Dashboard and read Needs attention: each row names what needs a look and links to its evidence. Or open Agents and sort the table by Success, lowest first. A success rate under 50% is Failing; 50% to under 80% is Degraded.
  2. Read the reason. Open the agent. The badge in the header carries the figure behind its status, for example “74.2% success among 31 labeled runs in the last 7 days”. The Success Rate card shows the change against the previous window.
  3. See the repeated failures. On the agent’s page or on Weekly Summary, open Repeated failures. Each group is runs that failed on the same kind of step with the same exception type and the same message (with ids and numbers removed). The “grouped on” line shows exactly why they landed together. Groups are ranked by what the failures cost, not by how many there are.
  4. Open the runs. Select View traces on a group. Or go to Traces, open More filters and set Success to Failed. Click a run to open it.
  5. Read the failing step. In the trace, find the step that failed or the last step before the run stopped. Compare it with a run of the same agent that succeeded.
  6. Check what changed. On the agent’s page, History shows whether a new version, revision, model or prompt shipped around when failures began. A failure that appears after a change is a lead to check, not proof the change caused it.
  7. If an alert fired, open Investigations. Morse’s first-pass hypothesis and its evidence are a starting point for your own check.

What the numbers mean

  • Success rate is successful runs divided by labeled runs. Runs with no outcome are left out, not counted as failures, so “74% of 31 labeled runs” tells you both the rate and how much it rests on.
  • A run can finish and still count as failed. A run’s status (completed, failed, timed out) says whether it ran to the end. Its outcome says whether the work held up. The Traces filters treat them separately: Status is on the toolbar, Success is under More filters.
  • Repeated failures describe what failed together. They do not say why.
  • New on a group means it first appeared in this window. Seen again after deploy describes timing only.

When there is not enough data

  • An agent with no labeled runs shows Inactive and no success rate. Morse will not call it healthy or failing without outcomes. Set up outcomes to fill it.
  • A failed run with no recorded exception is grouped by the step it ended on. Runs with no outcome label are never counted as failed.
Last updated on
QuickstartWhat does a good answer cost?