# Agent Builder

> Where you create an agent and tune everything about it — identity, prompt, memory, skills, connectors, built-in tools, channels, and storage.

**Agent Builder** is where an agent is created and everything about it is tuned. Every setting that changes how an agent behaves lives here, grouped into four sections.

## When to use it

Open Agent Builder when you want to change *what an agent is*, rather than ask it to do something. Creating an agent, rewriting its prompt, giving it a connector, turning a capability off, checking what it remembered — all of that happens here. Day-to-day work happens in a [conversation](/docs/work/conversations/) instead.

## Create an agent

<Steps>

1. **Start a new agent.**

   From your agent list, choose to create a new agent. You are asked for a name and a handle. The handle is short and stable — it is how you address the agent later.

2. **Say what it is for, in one sentence.**

   Be specific. *"Summarizes my unread email each morning and flags anything that needs a reply"* produces a far better agent than *"email assistant"*. This sentence is the seed of the [system prompt](/docs/build/system-prompt/).

3. **Open Agent Builder and work down the sections.**

   A new agent can already reason, calculate, and read public web pages. Everything else you add deliberately.

</Steps>

## The four sections

Agent Builder is organized by what part of the agent you are changing.

| Section | Contains | Page |
| --- | --- | --- |
| **Identity** | Profile, prompt, memory | [Identity and profile](/docs/build/identity/) · [System prompt](/docs/build/system-prompt/) · [Memory](/docs/capabilities/memory/) |
| **Ability** | Skills, connectors, built-in tools | [Skills](/docs/capabilities/skills/) · [Connectors](/docs/connect/overview/) · [Built-in tools](/docs/capabilities/built-in-tools/) |
| **Pairing** | Channels, devices | [Channels](/docs/automate/channels/) |
| **Inventory** | Workspaces, artifacts | [Artifacts](/docs/capabilities/artifacts/) |

The split is worth internalizing: **Identity** is who the agent is, **Ability** is what it can reach, **Pairing** is where it can be reached from, and **Inventory** is what it has produced.

## The anatomy of an agent

Three things determine how an agent behaves, in roughly this order of impact:

**Its system prompt.** The standing instructions it reads before every conversation. This is the single biggest lever, and the one most people under-use. See [System prompt and revisions](/docs/build/system-prompt/).

**What is bound to it.** An agent can only use the skills and connections you bind to it. Binding is per agent, not per account — connecting Gmail once does not hand your inbox to every agent you own.

**Its model.** Which model runs the loop, and how much reasoning effort it spends. See [Choosing a model](/docs/build/models/).

<Aside type="tip" title="Narrow beats broad">
  An agent with one job and three tools outperforms one with a vague purpose and
  twenty. When you find yourself widening an agent's prompt to cover a second job,
  build a second agent instead.
</Aside>

## Limits

- An agent's handle is chosen at creation and is not meant to change afterwards.
- Changes to an agent apply to its **next** run. A run already in flight keeps the configuration it started with — including the prompt revision, the model, and the bound tools.
- Agent Builder configures a single agent. There is no bulk edit across agents.
- Deleting an agent does not delete the [artifacts](/docs/capabilities/artifacts/) it produced.

## Related

<CardGrid>
  <LinkCard
    title="Identity and profile"
    href="/docs/build/identity/"
    description="Name, handle, profile, and the character that shapes its tone."
  />
  <LinkCard
    title="System prompt and revisions"
    href="/docs/build/system-prompt/"
    description="The biggest lever you have, and how to change it safely."
  />
  <LinkCard
    title="Choosing a model"
    href="/docs/build/models/"
    description="The model catalog, reasoning effort, and when each matters."
  />
  <LinkCard
    title="Testing and iterating"
    href="/docs/build/iterating/"
    description="Change an agent without breaking one that already works."
  />
</CardGrid>