> ## Documentation Index
> Fetch the complete documentation index at: https://docs.uselemma.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Monitors
> Track an event or behavior on an agent over time.
Use a **monitor** when you want continuous information about an event or behavior in an agent’s traces. It records matching traces as they arrive. See [Concepts](/platform/concepts#monitor) for the product definition.
The description tells the monitor what to look for. Each match updates its count and percentage. A behavior monitor also writes an issue when the agent gets that behavior wrong.
Use a monitor to track:
* How many times a user gets frustrated
* How many times an agent uses a tool wrong
* How many times an agent incorrectly cancels a reservation
Open **Monitors** in a project in the [Lemma dashboard](https://platform.uselemma.ai).
Use **All agents** or select an agent to filter the monitor list.
## Signal or behavior
You choose the kind when you create the monitor. It does not change later. Pick the kind by what you want back:
* **Signal**: records when an event occurs, with no issues. User frustration is a signal.
* **Behavior**: records when a behavior is wrong and writes an issue. A wrong tool call or cancellation is a behavior.
* **Template**: one monitor for each starter you pick
Those issues are separate from failures Lemma groups on its own. Alerting still follows the [issue lifecycle](/platform/issues).
## Create a monitor
**New monitor** asks for these, in order:
1. **Agent to watch.** Pick an agent for this monitor to judge. You can’t change it later. Lemma does not judge a trace that has no agent name.
2. **Kind.** Choose **Signal**, **Behavior**, or **Template**.
3. **Name.** This is the label on the card. Lemma does not look for a tool call with that name.
4. **Scenario**, on a behavior monitor. Describe what the agent should do.
For wrong tool use, name the request, the right action, and the mistake:
* Good: “A guest asks for their booking details. The agent looks up that booking, and no other.”
* Bad: “Use reservation tools correctly.”
On a signal monitor, **Which traces count** is optional. Use it to say what to count and what to skip.
## Percent and count
Each matching trace updates the count. The chart draws one point per day.
**Percent** is the share of that day’s traces that fit. **Count** is how many traces fit that day. A behavior monitor also shows how many of those traces have an occurrence.
You can view the same chart by week or by month. Until the monitor has seen a trace, there is no percent yet.
## When a trace is wrong
Tell Lemma when a trace is wrong. The chart and trace list update from the correction.
**Doesn’t fit** and **Mark incorrect** mean this trace should not count. **Add to monitor** means it should, including a trace the monitor has not judged. You can add a short reason. On **Traces** and **Issues**, you can mark several rows at once.
**Undo**, or **Restore** on **Feedback**, returns the trace to the monitor’s own judgment.
Select **Recalibrate** to teach the monitor from that feedback for later traces. Lemma also recalibrates every night.
The [correction rules](/reference/monitor-contract#corrections-and-counts), [learning window](/reference/monitor-contract#what-learning-keeps), and [recalibrate rules](/reference/monitor-contract#recalibrate) are in the monitor contract.
## Groups
**All monitors** lists every monitor in the project. **New group** adds a tab you name, and a monitor can belong to more than one group. **Existing** appears only inside a group, and adds monitors you already have.
Right-click a group to rename it, add monitors, or delete it.
Deleting a group leaves its monitors running. Everyone in the project sees the same groups and the same card order. Card order, creating on an open group tab, and the side rail are in the [monitor contract](/reference/monitor-contract#groups-and-the-board).
## From the API
REST and the [Lemma Model Context Protocol (MCP) server](/connections/mcp) share the reads. You create a monitor, change it, record feedback, and recalibrate on REST. What those calls change is in the [monitor contract](/reference/monitor-contract#from-the-api). Request shapes are in the [API Reference](/api-reference).
## Related pages
Failures Lemma groups on its own, and how alerting follows them.
The executions a monitor can judge.
How a project, an agent, and a trace fit together.
Read monitors from a coding agent. Writes stay on REST.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.