# Signals

> Define the behavior you want observation reviews to detect.

A Signal is a criterion Meta Agent checks when reviewing agent activity. Use Signals for behavior that matters to your application: an unsupported answer, an unnecessary escalation, a prohibited tool action, or a successful outcome you want to track.

Signals belong to the project. Review scope determines which related activity is considered together; the Signal describes what to look for within that activity.

## Define a criterion

Open **Observations → Signals**, choose **New signal**, and enter a title and a criterion under **What should Meta Agent look for?**.

Describe what constitutes a match and any exclusions that change its meaning. For example:

> **Escalation without a lookup** — Match when the agent escalates an order-status request without attempting the available order lookup. Exclude requests where the customer explicitly asks for a human.

The criterion should be assessable from the recorded conversation and tool activity. If an exclusion depends on customer input, that input needs to be present in the evidence reviewed.

Leave **Activate this signal** selected to use it in subsequent reviews, or clear it to save a draft. Choose **Create signal** to save it.

## Check the results

Record representative matching and nonmatching activity after activating the Signal, then [request a review](/docs/platform/observations/configure-reviews.md#review-new-activity). In the resulting [findings](/docs/platform/observations/review-findings.md), compare the verdict and rationale with the supporting records to check that the criterion reflects your intent.

If the matches do not reflect your intent, create a revised Signal and deactivate the old one. This keeps earlier verdicts attached to the criterion that produced them. Improving the wording cannot compensate for evidence your application never records.

## Manage active Signals

Use **Activate** or **Deactivate** on the Signal’s card:

* **Draft** saves a criterion without using it in reviews.
* **Active** includes it in subsequent reviews.
* **Inactive** stops future checks while retaining the definition and earlier verdicts.

Review suggested Signals before activating them, especially when an existing Signal already covers the behavior. Project owners, admins, and members can manage Signals; plan capacity can limit activation. See [Usage and billing](/docs/platform/administration/usage-and-billing.md).