Skip to main content
Not every task should run at the same level of independence. Sending a draft email for review requires different trust than running a web scrape. The autonomy dial gives you a single, continuous control — from zero to one hundred — that determines exactly how much authority your agents have before they stop and ask you.

​
The Autonomy Scale

The dial is divided into four behavioral zones. As you move the slider right, agents take on more responsibility and interrupt you less often.

0 – 20: Copilot Mode

The agent proposes every action and waits for your explicit confirmation before executing it. Nothing happens without your approval. Best for learning what agents do or working through sensitive workflows step by step.

20 – 50: Supervised

The agent executes low-risk actions automatically — reading data, browsing, drafting content — but pauses and asks before taking any action it classifies as high-risk, such as sending a message or modifying a file.

50 – 80: Autonomous

The agent handles the vast majority of actions independently, only interrupting for genuinely critical decisions: final sends, purchases, code deployments, or anything with irreversible consequences.

80 – 100: Full Auto

The agent executes end-to-end without interruption and notifies you after completion. Approval gates are bypassed. Use this level when you’ve validated the agent’s judgment and want maximum throughput.

​
Setting Autonomy

In the War Room UI: Use the autonomy slider in the War Room header. The setting takes effect immediately — you can adjust it at any point during a running task without pausing or resubmitting. When creating a task via API: Pass the autonomy_level field in your request body. 
curl -X POST https://api.miteos.com/v1/tasks \
  -H "Authorization: Bearer mt_live_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Research and draft outreach emails for our top 10 prospects",
    "autonomy_level": 40,
    "workspace_id": "ws_abc123"
  }'
The value must be an integer between 0 and 100. If omitted, it defaults to 70.

​
Approval Gates

Certain actions always trigger an approval gate when the autonomy level is below 80, regardless of context. These are actions with real-world consequences that cannot be undone. Actions that trigger approval requests:
  • Sending emails — any email.send call (drafts are always free)
  • Making purchases — commerce and checkout actions via virtual card
  • Executing trades — any live trade submission (backtesting never requires approval)
  • Deploying code — pushing to production or triggering a deployment pipeline
  • Deleting files — permanent deletion from the workspace or an external service
When an approval gate fires, the task pauses and a notification appears in the War Room Timeline as an approval.required event. You’ll see the exact action the agent wants to take, its parameters, and the options to Approve, Reject, or Edit the action before it proceeds. Via API, approval events are delivered as webhooks to your configured endpoint: 
{
  "event": "approval.required",
  "task_id": "task_xyz",
  "agent_id": "agent_email_1",
  "action": "email.send",
  "payload": {
    "to": "ceo@prospect.com",
    "subject": "Quick question about your data stack",
    "body": "..."
  },
  "approve_url": "https://api.miteos.com/v1/approvals/apr_123/approve",
  "reject_url": "https://api.miteos.com/v1/approvals/apr_123/reject"
}

​
Takeover Mode

At any autonomy level, you can take direct control of the agent’s browser without cancelling the task. Click Take Control in the Browser tab of the War Room Workspace panel. The agent pauses, your cursor takes over the live browser session, and you can navigate, fill forms, or complete any step manually. Click Return Control when you’re done, and the agent resumes from your current state. Takeover is particularly useful when a site requires a CAPTCHA, two-factor authentication, or a judgment call that only you can make.

​
Choosing the Right Level

Recommendations by use case:
  • First run on a new task type — Start at 20–40 to see exactly what actions the agent takes before granting more independence.
  • Research and analysis tasks — 60–70 works well; these tasks are mostly read-only and low-risk.
  • Content drafting — 50–70 lets agents write freely; approval gates will still catch anything that gets sent externally.
  • Outreach campaigns — 30–50 so you can review each email before it goes out.
  • Trusted recurring workflows — 80–100 once you’ve validated the agent’s behavior across several successful runs.
For trading and purchasing agents, start at 20 or below until you’ve validated the agent’s judgment across multiple paper-trade or dry-run sessions. Live trades and purchases at high autonomy levels execute immediately and may not be reversible.