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

# Metrics

> Executions, outcomes, tokens, cost and latency across the environment and per agent.

AgentOS counts everything that passes through it, so there is nothing to instrument in your agent. Every run, hold and model call made through the gateway is counted.

## Dashboard

The **Dashboard** is the environment's front page: executions, tokens, cost and outcomes across every agent, and a card per live agent (the denser its flow line, the busier the agent).

Pick a range in the top right:

| Range | Covers | Each bar is |
| - | - | - |
| **D** | the last 24 hours | an hour |
| **M** | the last 30 days | a day |
| **Y** | the last 12 months | a month |

What it shows:

* **Executions.** Every run counts, and a child run counts as an execution of its own. The throughput chart stacks each bucket's failed runs (red) on top of the others (green), on the same scale.
* **How executions ended.**

  * **completed on their own**: finished without asking anyone,
  * **went to a person**: asked a person at least once, whatever happened next,
  * **failed or cancelled**.

  A held run is not a failure: its agent asked a person to decide.
* **Waiting for a person now**: the holds open at this moment. Decide them in the [Review queue](https://docs.upsonic.ai/agentos/platform/review-queue.md).
* **Tokens** and **Cost** of the model calls. Costs come from the gateway's price table for each model.
* Changes against the previous period of the same length.

## An agent's Overview

The agent's **Overview** tab shows the same numbers for one agent, with the same D / M / Y ranges:

* **Executions** (and their rate per hour, day or month), **Tokens** (per run), **Cost**, and **Held for a person**, with how long the oldest one has been waiting.
* **Executions and latency**: throughput, with p50 and p95 run duration. A run's duration includes the time it waited for a review. The latency line breaks where a bucket had no runs.
* **How runs ended**, as on the Dashboard.
* **What went wrong**: failed runs grouped by error code. Guardrail blocks are also grouped by their message.
* **Recent runs**, **Versions**, and the **Declared configuration** and **Capabilities** of the selected version, from its `agentos.yaml`.

## Models page usage

The **Models** page shows each model's calls, tokens and cost over the last 30 days. A model that cost nothing is marked **on-prem**. See [Models and providers](https://docs.upsonic.ai/agentos/platform/models-and-providers.md).

## How the numbers are counted

* **Cost.** A model the price table doesn't know costs $0. A cost under a cent shows as `<$0.01\`. "On-prem models cost \$0" is shown only when every model called in the range cost nothing.
* **Eval runs** are not counted anywhere on these pages, and an AI judge's tokens appear on its experiment, not on the Models page. See [Evals](https://docs.upsonic.ai/agentos/platform/evals.md).
* **Times are UTC.**

For one run's own tokens, cost and model calls, open the run: its `usage` and timeline list every call. See [Runs API](https://docs.upsonic.ai/agentos/api/runs.md).


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