Your agent can fail while every trace stays green because most agent failures aren't Exceptions. An Exception is a call that failed: an error, a 4xx or 5xx response, a timeout. A Misstep is a call that ran successfully but did the wrong thing, such as calling the wrong tool or passing the wrong parameter. Status codes only see Exceptions. To see Missteps you need a label on every span that says what the call did and whether it worked.
A status code is enough for application logs, because the same request runs the same code every time. An agent is nondeterministic. The same request can pick a different tool or a different argument on the next run, and a call can succeed while doing the wrong thing. To judge a span you have to read its input and its output together.
Kubit does that with Polaris, its in-house small language model, in two ways. Label with Classifier puts an annotation type such as Misstep or Exception on every span. Score with Eval scores every span against questions you write, such as "Did the agent book the date the user asked for?" Both can be fine-tuned on the corrections your team makes.
Key facts
Exception: the call failed, through a thrown error, a 4xx or 5xx response, a timeout or no network (Concepts)
Misstep: the call ran successfully but did the wrong thing, for example
wrong_toolorwrong_parameter(Concepts)Where each one shows: an Exception shows in your error logs. A Misstep usually doesn't (Concepts)
How Kubit finds them: Polaris labels and scores every span, not a sample, with no extra LLM cost (Polaris)
What is an Exception in an agent trace?
An Exception is a span where execution failed. The tool threw an error, the API returned a 4xx or 5xx, the call timed out or the network dropped. Labels such as time_out and unauthorized fall here. Exceptions are the failures your existing monitoring already catches, because the trace records them as errors.
What is a Misstep?
A Misstep is a span where the call ran successfully but did the wrong thing. The model picked search_orders when it needed get_refund_status, or it passed last Tuesday's date to a booking tool when the user asked for next Tuesday. The tool did exactly what it was told, so it returned OK. The trace stays green. The user still got the wrong result.
Why status-code monitoring misses Missteps
A status code tells you whether a call completed. It says nothing about whether the call was the right one. An agent decides which tool to call and with which arguments, so the most common way it goes wrong is a correct call to the wrong thing. That shows up later as a re-prompt, a support ticket or a user who leaves, and none of those sit in the trace as an error.
One trace, two views

Take an illustrative booking agent. The user asks for an appointment next Tuesday at 3pm.
Span | Status | Polaris label |
|---|---|---|
| OK | Action |
| OK | Action |
| OK | Warning: |
| OK | Misstep: |
| Error | Exception: |
A status dashboard shows one problem: the confirmation timed out. Polaris shows three. The repeated search is a Warning, the wrong date is a Misstep, and the timeout is an Exception. The Misstep is the one that cost the user an appointment, and it's the only one with an OK status.
How do you detect Missteps in your agent traces?
You label every span with what it did and whether it worked, then look for Misstep labels on traces whose status is OK. In Kubit, Polaris labels and scores every span, and the Labels column in Activity shows each trace's labels, such as wrong_tool or wrong_parameter. Open a trace and the timeline shows each span's input and output, so you can check what the agent called and with which arguments. The Find Missteps that return OK guide walks through it in about five minutes once traces are labeled.

Is it one trace or a pattern?
A single Misstep is worth a ticket. A Misstep on a path many runs take needs a fix in the agent. Open Flows, which visually groups every run by execution path (which tools ran, in what order) and shows how many traces took each path. If the wrong_parameter on book_appointment sits on a path that a fifth of your runs take, you know where to look first. Then compare the Misstep rate across model versions to see whether a release caused it.
Why every span, and not a sample
Missteps are quiet by definition, so a sample will miss many of them, and a path that only some runs take can vanish from a sampled view altogether. Polaris labels and scores every span because Flows needs every run labeled to show every execution path with its count. We wrote about that choice in Why we label every span instead of a sample.
Where to start
If your traces look healthy and users still complain, start by measuring Missteps. Join the early access list, or book a time with Alex to go through your own traces.
Read next: Your agent traces are green and it's still failing · Why we group agent runs by execution path · What is Kubit?
FAQ
What is the difference between a Misstep and an Exception?
An Exception is a call that failed: a thrown error, a 4xx or 5xx response, a timeout or no network. A Misstep is a call that ran successfully but did the wrong thing, such as calling the wrong tool or passing the wrong parameter. Exceptions show in error logs. Missteps usually don't.
Why is my AI agent failing when its traces are green?
Most likely because of Missteps. The agent called a tool that worked, so the span returned OK, but it was the wrong tool or the wrong arguments. Status codes only record Exceptions, so the trace stays green while the user gets the wrong result.
How do I detect silent failures in an LLM agent?
Label every span with what it did and whether it worked, then look for Misstep labels such as wrong_tool or wrong_parameter on traces whose status is OK. In Kubit, Polaris labels and scores every span and Activity shows each trace's labels.
Does Kubit sample traces to find Missteps?
No. Polaris labels and scores every span, not a sample, with no extra LLM cost. Flows needs every run labeled to show every execution path with its count.
Keep reading

What is Kubit? Agent Analytics, from trace to user

Langfuse alternatives for agent analytics in 2026

Braintrust alternatives for agent analytics in 2026

Agent analytics on traces already in Snowflake, Databricks or BigQuery

Should you build your own agent analytics?

