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

# Training

> Review an agent's work, and send it away to study

You train an agent two ways: by **reviewing its work before it acts**, and by **sending it away to study** something too large for one conversation. Both write back to the same [memory](/docs/agents/memory), so what it learns sticks.

## Reviewing before it acts

Training mode lets you check an agent's output before it goes out. Use it while you build trust in a new agent, dial in a workflow, or work on something where a mistake would be costly.

In training mode, the agent presents its work as **artifacts** for review instead of sending or executing directly. You see exactly what it intends to do. You approve or revise it. Only then does it go out.

### When to use it

<CardGroup cols={2}>
  <Card title="A new agent" icon="seedling">
    Watch how it handles real tasks before you let it act unsupervised.
  </Card>

  <Card title="A new workflow" icon="route">
    Validate a fresh skill or process end to end before trusting it.
  </Card>

  <Card title="High-stakes output" icon="triangle-exclamation">
    Customer-facing replies, anything irreversible, anything you want a human gate on.
  </Card>

  <Card title="Tuning the persona" icon="sliders">
    See how persona and instruction changes land before they reach anyone.
  </Card>
</CardGroup>

### The flow

Turn it on by asking, in plain language:

```text theme={null}
Go into training mode. I want to review your replies before they go
out while we dial in the support tone.
```

<Steps>
  <Step title="Turn on training mode">
    Ask the agent to enter training mode in your conversation. It applies to that agent until you turn it back off.
  </Step>

  <Step title="The agent drafts">
    It does the work and presents the result as an artifact instead of acting.
  </Step>

  <Step title="You review">
    Approve it, edit it, or send it back with feedback.
  </Step>

  <Step title="Graduate">
    Once the agent consistently gets it right, turn training mode off and let it run on its own.
  </Step>
</Steps>

<Tip>
  Pair training mode with the [feed](/docs/workspaces/feed): review what the agent produced in the moment, then audit the trail afterward.
</Tip>

## Sending it away to study

Some things take real study: your communication style, the shape of a project, your key relationships, a body of reference material. A normal conversation is too small a window. In a **deep learn** session the agent studies in the background and comes back having absorbed it.

<CardGroup cols={2}>
  <Card title="Learn your style" icon="pen-nib">
    How you write and communicate, so its output sounds like you.
  </Card>

  <Card title="Learn a project" icon="diagram-project">
    The context, decisions, and players behind a body of work.
  </Card>

  <Card title="Learn your world" icon="people-arrows">
    Key relationships, accounts, and the things you care about most.
  </Card>

  <Card title="Learn a domain" icon="book">
    Reference material and documentation it should master.
  </Card>
</CardGroup>

### How a session starts

A session begins one of two ways:

* **You ask.** Tell the agent to get up to speed on something: "spend some time learning our brand voice," "get to know my top accounts."
* **The agent proposes one.** When a task needs far more context than a chat holds, the agent suggests a session.

Either way, **you confirm before it runs**. A long session never starts without your go-ahead.

Once confirmed, the session runs in the background. You do not have to stay in the conversation. The agent works through the material on its own and reports back with a summary of what it learned.

### Scoping a session

A good session starts with a clear goal. Before it begins, the agent works out with you:

* **What to focus on**: the specific area to study.
* **What "good" looks like**: concrete outcomes ("you should know my top five accounts," "match my email tone").
* **What to study**: the people, projects, channels, or docs that matter.
* **What it should handle afterward**: the tasks the learning should enable.

With the scope set, the agent runs the session and writes what it learns to its [memory](/docs/agents/memory). The new context applies to every future conversation.

<Tip>
  Deep learn pays off most for an agent you will rely on long-term, not a one-off task.
</Tip>

<CardGroup cols={2}>
  <Card title="Where the learning lives" icon="brain" href="/docs/agents/memory">
    Everything training and study produce is written to durable memory.
  </Card>

  <Card title="Shape behavior as you review" icon="masks-theater" href="/docs/agents/persona">
    Fold what you learn in training back into the agent's persona.
  </Card>
</CardGroup>