When you set a tab to Manual, Auto-Supervised, or Fully Autonomous, the agents don't run on a different program for each mode. Instead, each agent is handed the rules for the current mode as a message — the way you'd brief a coworker on how to handle this shift. This is a look under the hood at how that works, and why the mode selector behaves the way it does.
If you just want to pick a mode, start with AI modes. This page is for understanding what's happening beneath it.
Rules as a message, not a rebuild
The mode-specific rules are injected as a dispatcher-style message, not baked into the agent's underlying prompt. In practice that means the agent gets a note that reads like an instruction from you — it opens with "[SYSTEM_NOTICE] Operating mode is now …" — telling it how to behave for the mode the tab is in.
The upshot for you: switching modes doesn't rebuild the agent, it re-briefs it. The same agent follows tighter or looser rules depending on the note it's been given.
Per agent, per mode
Each agent gets its own rules for each mode — the chat agent, the email agent, and the broker-outreach agent each have their own instructions for Manual, Auto-Supervised, and Fully Autonomous. So a reply drafted by the email agent in Fully Autonomous follows different guidance than the same agent would in Manual, and the broker-outreach agent has its own set again.
This is why the negotiation and approval behavior you read about elsewhere shifts with the mode: it's the injected rules for that agent, in that mode, doing the shifting.
Why the selector locks mid-run
Because the rules ride in as a message at the start of a run, changing modes in the middle of one would mean the agent is partway through a job under one set of instructions when the rules underneath it change. To avoid that, the AI-mode selector is disabled while a run is in flight — if you try, you'll see a tooltip telling you to wait for the current run to finish. Let it wrap up, then switch, and the next run starts cleanly under the new mode's rules.
Related
- AI modes — Manual, Auto-Supervised, and Fully Autonomous compared.
- Fully Autonomous mode and personas — where a persona playbook can override the default behavior.
- How the send classifier works — the other half of how sending is governed per mode.
Need help? Email support@numeo.ai or use Report a problem in the app.