# What agents do

An <Term id="agent">agent</Term> works on its own. It reads your data, does something with it (runs code, asks an AI, researches the web) and saves the result to a table, a file or a short answer. You describe the job in plain words, and the AI in your project's [Studio](/docs/projects/studio) builds it as a draft.

Use one for work that repeats: a morning summary, scoring new rows, copying a web API into a table every hour, a weekly report.

## How an agent is built [#how-an-agent-is-built]

Every agent belongs to one [project](/docs/projects/create-a-project) and is made of **steps**. Each step does one thing, and later steps use what earlier ones made. Click any step on the agent's page to see its settings.

| Step | What it does |
| --- | --- |
| **Code** | Runs a short Python file you or the AI wrote. |
| **AI** | Asks an AI model to read the step's input and answer an instruction. |
| **Researches the web** | Looks things up on the web and writes down what it found. |
| **Reads data** | Runs a <Term id="sql">SQL</Term> query on [Nexus](/docs/data), your database. |
| **Saves to a table** | Writes rows to a <Term id="table">table</Term>: **Add rows**, **Update or add**, or **Replace all**. |
| **Asks you first** | Pauses the run and waits for a person to say yes or no. |
| **Starts another agent** | Hands work to another agent. |

A step can run once **for each item** an earlier step made, with a setting for how many run **at once**. It can run **only when** a condition is true, retry after a failure (**Retries**) and stop after a **Time limit**. An AI step can use a specific **Model**; left on the plan default, it follows your plan.

## When an agent starts [#when-an-agent-starts]

- **On a schedule**: every hour, daily at a set time, or on chosen days of the week.
- **When files arrive** in a folder in [Object Storage](/docs/data/object-storage).
- **When rows load** into a [table](/docs/data/tables).
- **When a webhook calls**.
- **When you press Run**.

Automatic starts only work once the agent is live. Until then it is a draft you can try as often as you like.

## How a run works [#how-a-run-works]

Each start makes a <Term id="run">run</Term>. It goes through the steps in order, and each step shows its state: Waiting, Running, Done, Failed, Skipped, or **Waiting for you** when it asks for approval. A run ends as Done, Failed, Declined (someone said no at an approval) or Stopped.

The run's **Result** shows what the agent produced: rows, text or files to download. Every step keeps its log and files. See [Runs and approvals](/docs/agents/runs).

## Agents ask before changing anything important [#agents-ask-before-changing-anything-important]

An **Asks you first** step can go before anything that saves or changes data. The run pauses there and the agent appears under **Needs you** on the Agents page. Nothing after it happens until a person chooses **Yes, continue**.

## Drafts, live agents and who decides [#drafts-live-agents-and-who-decides]

You work on a **draft**. **Try it** runs it once, now, and nothing goes live. When it does what you want, a <Term id="workspace">workspace</Term> admin chooses **Go live**. From then on it starts on its own and spends the workspace's credits. Any earlier version can be made live again. See [Try, go live and pause](/docs/agents/go-live).

<Callout type="info" title="Only a person takes an agent live">
  The AI in your Studio can build and try agents, but it cannot take one live. When it asks to publish an agent, you see the request and decide. Only a workspace owner or admin can take an agent live.
</Callout>

## What agents cost [#what-agents-cost]

Runs use <Term id="credits">credits</Term> from your workspace. Each run's cost is in the **Credits** column of the runs list, and the Agents page shows **Credits today** and runs for the last week. Balance and spend are in <AppLink to="settings.usage" />.

## The Agents page [#the-agents-page]

<AppLink to="agents" /> shows every agent in the projects you can open, grouped by what needs attention first:

- **Needs you**: a run is waiting for your approval.
- **Running**: a run is in progress.
- **Failing**: the latest run failed.
- **Automatic**: live and starts on its own.
- **On demand**: live and waits to be started.
- **Not live yet**: still a draft.

Switch between **Board** and **List**, or search by agent or project name.

<Screenshot src="/docs/_media/agents-overview.png" alt="The Agents page with agents grouped by what they need" />

## Where to go next [#where-to-go-next]

- [Create an agent](/docs/agents/create-an-agent)
- [Templates](/docs/agents/templates)
- [Try, go live and pause](/docs/agents/go-live)
- [Runs and approvals](/docs/agents/runs)