> ## Documentation Index
> Fetch the complete documentation index at: https://learn.kubit.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Find Missteps that return OK

> Find agent calls that returned an OK status but did the wrong thing, using Polaris Misstep labels in Activity or from an AI assistant.

A Misstep is a call that ran, often with an OK status, but did the wrong thing. Polaris labels Missteps on every span, so a trace whose calls all return OK can still show a `wrong_tool` or `malformed_args` label. This page shows how to find those traces in Activity or from an AI assistant. It takes about 5 minutes once your traces are labeled.

## Before you start

* [ ] A workspace with agent traces. See the [Quickstart](https://learn.kubit.ai/agent-observability/get-started/quickstart.md).
* [ ] Traces labeled by Polaris. In **Activity**, the **Labels** column reads `Not analysed yet` until a trace is labeled.

## Steps

### 1. Open Activity

Open **Agent** > **Activity**. Keep **Subject** on **Trace**, set **Time Range** and click **Execute**.

### 2. Read the Labels column

The **Labels** column shows each trace's labels. Look for Misstep labels such as `wrong_tool`, `malformed_args` or `session_not_completed` on traces whose status is OK. Warning labels such as `repeated_tool_call` are worth a look too.

![The Activity traces grid with a Timeline of trace counts in 30-minute bars and errors in red, above the columns Timestamp, Name, Spans, Latency (Ms), Tokens, Model, Labels, Trace ID and User ID. The Labels column shows labels such as completed, session\_not\_completed, answered and low\_judge\_confidence. The footer reads 519 Traces, 1-50 shown.](https://mintcdn.com/kubit/iy-nIyF9Xtsr91WV/images/agent-observability/activity-labels.png?fit=max\&auto=format\&n=iy-nIyF9Xtsr91WV\&q=85\&s=363d5e7136da16ebb01235a5b51ed41a)

The **Labels** column can't be filtered or sorted. On a Sandbox or warehouse workspace, filter by label instead: click **+** next to **Filter**, choose **Filter** and pick **Misstep Label** (or **Evaluator: Misstep Label**), then a value such as `wrong_tool`. **Trace Misstep Label** matches the trace's label rather than a span's.

### 3. Open a trace

Click a trace to open it. The **Timeline** shows its agent runs, model calls and tool calls. Click a span to see its **Model**, token counts and **Input** and **Output** messages, and check what the agent called and with which arguments.

![An open trace named classify\_intent with an Error status, Latency 451ms, Tokens 341 to 51, a User and a Session. The Timeline shows a retrieve\_products span tagged not\_found. The Details panel's Preview tab shows the Input messages. Below, Labels and feedback lists the Evaluator questions, each answered No with a score.](https://mintcdn.com/kubit/iy-nIyF9Xtsr91WV/images/agent-observability/trace-span-label.png?fit=max\&auto=format\&n=iy-nIyF9Xtsr91WV\&q=85\&s=92647a270cec82bc4e549c53b144c5f4)

**Labels and feedback**, at the bottom of the open trace, lists **Kubit's labels**. Each label shows its scope, **Whole trace** or the span it labels, and buttons to confirm or reject it. Your verdicts tune that label's threshold, and a rejected label no longer shows in the **Labels** column. In the **Timeline**, a failed span is tagged with its exception kind, such as `not_found`.

### 4. See how often it happens

Open [Flows](https://learn.kubit.ai/agent-observability/use/flows.md) to see whether the Misstep is one trace or a path many traces take. Each path shows how many traces took it.

## Build with a prompt

Ask Lumos or an AI assistant with the Kubit MCP server connected. The `/kubit-inspect` skill returns the total number of matches, up to 5 sample rows and a CSV export of every match.

> "Show me traces from yesterday with a Misstep label"

> "Which tool calls had a Misstep label in the last 7 days, grouped by model version?"

## Verify

The traces you found have an OK status and at least one Misstep label. Open one and confirm the labeled span called the wrong tool or passed the wrong parameter.

## If it fails

| **Symptom** | **Cause** | **Fix** |
| - | - | - |
| **Labels** reads `Not analysed yet` | Polaris hasn't labeled the trace | On an Agent workspace, Kubit labels new conversations every hour. On a Sandbox or warehouse workspace, start a run from the **Evaluator** page. Then click **Refresh**. |
| **Labels** reads `No labels` | Polaris labeled the trace and found nothing to label | Nothing is wrong. Open the trace to answer the Evaluator's questions yourself. |
| **Labels** is empty on every trace | The workspace has no label source | Check that the Evaluator is set up for the workspace. See [Polaris](https://learn.kubit.ai/agent-observability/use/polaris.md). |
| **User** is empty on every trace | The spans carry no user ID | Set `user.id` in your app. See [Add user and session IDs](https://learn.kubit.ai/agent-observability/integrate/sdk-python.md#add-user-and-session-ids). |

## Next Steps

* [Polaris](https://learn.kubit.ai/agent-observability/use/polaris.md): the classifier and the five annotation types.
* [Find what changed before a pattern started](https://learn.kubit.ai/agent-observability/use/find-what-changed.md): find the commit behind a rise in Missteps.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.