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

# The agent

> Chat with an agent that browses, scrapes, analyses and acts on your workspace.

The chat at the centre of the dashboard is a full agent, not a chatbot. It
browses the live web, runs your scrapers, reads spreadsheets, queries your
connected databases, saves leads, and asks you questions when it is unsure.

## Runs

Every message that requires work becomes a **run**:

```
queued → planning → running → completed
                  ↘ paused (waiting on you)
                  ↘ failed / cancelled
```

While a run is going you see its steps stream live. When it finishes you get a
result summary plus any **artifacts** — CSV files, datasets, documents — which
stay downloadable from the run.

Over the API, runs are
[`POST /api/agents/runs`](/api-reference/agent/create-run) and
[`GET /api/agents/runs/{id}`](/api-reference/agent/get-run).

## It asks questions

When your instruction is ambiguous, the agent **pauses and asks** instead of
guessing. That is a feature: a lead search that guesses wrongly about your ICP
spends real credits producing the wrong list.

In the dashboard, answer in the chat. Headlessly:

<Steps>
  <Step title="Poll the run">
    `GET /api/agents/runs/{id}` returns `status: "paused"` and the question in
    `result_summary`.
  </Step>

  <Step title="Answer it">
    `POST /api/agents/runs/{id}/reply` with `{ "message": "…" }` resumes the
    **same** run mid-thought — no new run, no lost context.
  </Step>
</Steps>

Replying to a run that has already finished starts a follow-up run in the same
conversation, inheriting its context and playbook. So *"find 25 more like
those"* keeps saving leads rather than starting from nothing.

Pass `"interrupt": true` to stop a still-working run and immediately start a new
one with your message.

## Cancelling

[`POST /api/agents/runs/{id}/cancel`](/api-reference/agent/cancel-run) stops a
run right away. Cancellation is authoritative, not advisory — the run is marked
cancelled immediately, so a wedged executor cannot keep it "running" forever and
occupying one of your parallel-run slots.

## What the agent can do

| Capability                     | Notes                                                                       |
| ------------------------------ | --------------------------------------------------------------------------- |
| Search and browse the web      | Costs crawl credits per page.                                               |
| Run your saved scrapers        | Structured, schema-validated extraction.                                    |
| Find and enrich leads          | The `lead_generation` playbook — see [Finding leads](/leads/finding-leads). |
| Analyse spreadsheets and files | Upload them or point at Vault artifacts.                                    |
| Query connected databases      | Read-only Postgres, MySQL or BigQuery — see [Data](/agent/data).            |
| Build charts                   | The visualization Studio (Pro and Business).                                |
| Call your MCP connectors       | Third-party MCP servers you have connected.                                 |

## Playbooks and sources

A run can be pointed at specific inputs and behaviours:

| Field                     | Meaning                                           |
| ------------------------- | ------------------------------------------------- |
| `goal`                    | What you want, in plain language.                 |
| `playbook`                | A named behaviour, e.g. `lead_generation`.        |
| `sources`                 | Files, scraper runs or URLs to ground the run in. |
| `output_schema`           | A JSON schema the result must conform to.         |
| `max_steps` / `max_pages` | Budget guards for a long run.                     |

## Conversations, folders and brands

Runs hang off **conversations**, which are scoped to your workspace and narrowed
to the brand you are working in. `?brand=all` widens a listing to every brand in
the workspace — useful for agencies — but never past the workspace.

Organise conversations into **folders** from the sidebar.

## Parallel runs

| Plan     | Runs at once |
| -------- | ------------ |
| Free     | 1            |
| Starter  | 2            |
| Pro      | 5            |
| Business | 10           |

Pro and Business also get the full-size model with configurable reasoning
effort; lower tiers run on the lighter models.

<Card title="Skills" icon="wand-magic-sparkles" href="/agent/skills" horizontal>
  Save an instruction set as a reusable, selectable playbook.
</Card>