Autonomy levels, chosen safely.

Every agent in Orbit runs at one of three autonomy levels, set by you and changeable at any time. Choosing well is not about trust in AI; it is about matching the level to how reversible the action is. This guide gives you a working rule for each mode and a promotion path that does not rely on optimism.

The three modes, honestly

• Manual. The agent acts only when you tell it to, like a very fast colleague waiting for instructions. Manual runs never cost a credit, which makes this the natural mode for testing and one-off jobs.
• Approve-each. The agent does the work on its own schedule, but every result waits in the Approvals queue for your yes. You review outcomes, not effort. This is the default for anything customer-facing, and the mode where most work should start.
• Auto-run. The agent acts and the log records it. You review after the fact. Reserve this for actions you have already approved dozens of times without edits.

What should never start on Auto-run

Some actions are cheap to undo; some are not. A drafted social post that waits a day costs you nothing. A wrong email to a customer, a payment reminder to someone who already paid, a discount that violates your own policy: those are expensive in the currency that matters, which is trust. The rule:

• Anything that moves or requests money: Approve-each until it has been boring for weeks.
• Anything a customer or client sees under your name: same rule.
• Anything irreversible outside the product, like a published post or a sent message: same rule.
• Internal drafts, research, enrichment, and categorization: promote these first; a wrong draft costs you a shrug.

Autonomy is per action, not per account

You do not choose between "trust the agent" and "do not." An outreach agent can source and score leads on Auto-run, draft messages on Auto-run, and send them on Approve-each. Split the pipeline at the point of no return and be generous on everything before it, strict on everything after.

The backstops

Two things make the promotion decision low-stakes. The run log records every action, timestamped and replayable, so "what did it actually do" is never a mystery. And the kill switch pauses every agent in the workspace with one tap, so the worst case of a bad promotion is a short pause and a demotion back to Approve-each, not a cleanup project.

In workflows the same grammar applies: Auto-run steps fire on their own, Approve-each steps hold the chain until you clear them. Build the chain so the pause lands exactly where the risk does.