Low-signal traces
The Traces list marks a run as low-signal when it recorded nothing that tells you how an agent behaved: no model call, no cost, no tokens. Once your workspace has enough real AI traces, low-signal runs are hidden from the list by default so they do not bury the runs that matter. You can show them at any time.
What counts as low-signal
A trace is low-signal only when all of these are true:
- It has no cost.
- It has no tokens.
- It lasted 100 ms or less.
- It has no graph or workflow metadata.
- It has no span other than a bare agent span.
- No span carries a model.
If any one of these is false, the trace is not low-signal.
Failed runs are never low-signal. A run with status failed or timed_out always shows in
the list, whatever else it did or did not record.
When they are hidden and when they show
Morse decides automatically, based on how many real AI traces the workspace has. A real AI trace is one that recorded a model call: tokens, a cost, or a model span. Failed runs do not count toward the total.
- Fewer than 10 real AI traces. Low-signal traces are shown, so your first run is never invisible.
- 10 or more. Low-signal traces are hidden by default. A one-time notice tells you so, and a row above the results stays visible: N low-signal traces hidden · Show them.
- Your own choice wins. Once you turn Show low-signal traces on or off, Morse remembers it and stops applying the automatic rule.
How to show them
Either of these works:
- Turn on Show low-signal traces in the Filters sheet.
- Click Show them in the count row above the results.
Why a run is low-signal
Common causes:
- A
run()around code that makes no model call. SDK setup, smoke tests and hello-world scripts record an agent span and nothing else. - Framework scaffolding. Some frameworks emit short chains of internal steps, for example LangGraph state updates, that carry no model call.
- A guardrail or early exit. An agent that checks its input and returns before calling a model is doing what it should. It is shown as low-signal because the rule reads only the signals above and cannot tell intent.
- A model client that is not wrapped. The model call happened, but nothing recorded it. See make a run count as real AI.
- A trace with zero spans. Nothing was recorded for the run. If this is unexpected, check that the SDK flushes before the process exits, which matters most in short scripts.
Why input and output may be empty
A bare run records no input or output unless your code sets them, so the trace detail can show No input recorded and No output recorded. Input and output appear when the run records them. See the SDK overview for what the SDK records.
Make a run count as real AI
Wrap your model client with the SDK adapter, and each model call is recorded with its model, tokens and cost. That is what makes a run count toward the 10 real AI traces and keeps it in the default list. The Frameworks page has the adapters for Anthropic and OpenAI, and the Quickstart shows a first run that makes a real model call.