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

# Human tasks

> Add approval gates and human-in-the-loop pauses to AgentRuntime workflows.

**Human tasks** let you pause a workflow until a person reviews, approves, rejects, or edits data. They are essential for production agents that take real-world actions — sending emails, updating CRM records, or processing payments.

## How human tasks work

1. A workflow reaches a `human_task` step
2. The run **pauses** and creates a pending task with the `task_payload` you defined
3. The task appears in **Command Center** and in per-run task lists
4. A project contributor **completes** the task (approve, reject, or submit edits)
5. The run **resumes** with the completion result available to downstream steps

## Authoring a human task step

Add a step with `type: "human_task"`:

| Field          | Description                                                                           |
| -------------- | ------------------------------------------------------------------------------------- |
| `task_type`    | Category such as `approval` or `edit`                                                 |
| `task_payload` | JSON shown in the UI — title, description, fields to review, optional editable values |

Keep payloads concise. Include enough context for a reviewer to decide without opening external systems.

## Completing tasks in the Console

See the [Command Center guide](/workflows/command-center) for the full operator inbox.

<Steps>
  <Step title="Open Command Center">
    Go to **Command Center** in the sidebar. Pending tasks are listed with workflow name, run ID, and task title.
  </Step>

  <Step title="Review the payload">
    Open the task to see the data the workflow paused on — draft email, proposed CRM update, generated summary, etc.
  </Step>

  <Step title="Approve or reject">
    Click **Approve** to continue the run with an approval result, or **Reject** to stop the branch with a rejection result. Some task types allow editing fields before submission.
  </Step>
</Steps>

You can also complete tasks from Workflow Studio's run panel when viewing an active run.

## What downstream steps receive

When a task completes, the `human_task` step result is a JSON object:

| Field      | Always present | Description                             |
| ---------- | -------------- | --------------------------------------- |
| `approved` | Yes            | `true` if approved, `false` if rejected |

The Console **Approve** and **Reject** buttons send only `{ "approved": true }` or `{ "approved": false }`. They do **not** rename or edit draft fields automatically.

To pass edited text to downstream steps, either:

* Reference the **upstream LLM/MCP step** that produced the draft (for example `{{steps.draft-review.result.text}}`) and gate with `condition: "{{steps.approve-review.result.approved}}"`
* Complete the task via API with an optional `result` object (merged into the step result alongside `approved`)

Example API body with edits:

```json theme={null}
{
  "approved": true,
  "result": {
    "comment_body": "Edited review text posted to GitHub."
  }
}
```

Downstream templates would use `{{steps.approve-review.result.comment_body}}`.

## API completion

Automate or integrate approvals via the API:

```
POST /v1/workflows/{id}/runs/{runID}/tasks/{task_id}/complete
```

Requires **project\_contributor** and workflow run scopes on your PAT.

List pending tasks:

* Per run: `GET /v1/workflows/{id}/runs/{runID}/tasks/pending`
* Tenant inbox: `GET /v1/tasks/pending`

## Design patterns

| Pattern                 | When to use                                                                              |
| ----------------------- | ---------------------------------------------------------------------------------------- |
| **Approve before send** | LLM drafts content; human approves before an MCP email/send step                         |
| **Escalation**          | Low-confidence LLM output routes to human review via a `condition` + `human_task` branch |
| **Data correction**     | Human edits extracted fields before downstream MCP writes                                |

<Warning>
  Runs paused on human tasks hold worker resources. Set timeouts on preceding steps and monitor Command Center so approvals do not stall production runs.
</Warning>