Choose an autonomy level for an agent

For any operator who wants to adjust how far an agent can go on its own before asking for a confirmation.

Prerequisites

• Apollia running and the daemon active.
• At least one agent installed.
• Familiarity with the apollia-os run command.

The four levels

Level	When to use it	Step budget	Automatic verification	Memory injection
assisted	Exploratory work, unknown task, first run.	Very short - suitable for validating a prototype.	None. The plan gate is armed, so you approve the plan before it runs.	No.
supervised	Standard daily use.	Moderate - covers most everyday tasks.	One verification pass after the run finishes.	No.
bounded_autonomous	Configured automations, recurring pipelines whose behavior you already know.	Generous - suited to long but bounded workflows.	One verification pass after the run finishes.	No.
long_autonomous	Long-running background tasks, overnight work on large volumes of data. Reserved for proven agents.	Maximum available.	One verification pass after the run finishes.	No, for an agent.

Two things the table would let you assume, and should not.

Approval is not per step. No level pauses before each action. What assisted and supervised arm is the plan gate: you see the plan and approve it once, before execution. bounded_autonomous and long_autonomous bypass that gate entirely, so the plan runs without you seeing it, unless you pass --plan to re-arm it for that run. Separately, and at every level, a filesystem write the runtime judges risky raises its own approval prompt.

Verification runs once, at the end. It is a single post-run pass, not a check between steps, and assisted does not run it at all.

The "Memory injection" column is No for every level on an agent run, and that is the guarantee that matters: nothing injects memory into an agent's prompt. The built-in conversational assistant is a different path: at every level it receives the profile of the conversation, while the "Apollia knows my profile" switch of that conversation is on. An agent cannot reach that code.

Steps - Apply a level to one run

The level is set at launch, with --autonomy. It applies to that run only and does not modify apollia.toml.

apollia-os run my-agent "my task" --autonomy supervised

Replace supervised with one of the four values: assisted, supervised, bounded_autonomous, long_autonomous.

Steps - Change the level for every run

There is no global default level to set. [autonomy] is not one of the nine sections the daemon reads, so a default_level written in apollia.toml changes nothing, and the command that would write it is refused:

$ apollia-os config set autonomy.default_level supervised
Error: unknown config key: 'autonomy.default_level'

A run launched without --autonomy takes the level of the agent, and assisted when the agent names none. To give one agent another level for every run, set it on that agent:

apollia-os agent set my-agent autonomy supervised

apollia-os agent reset my-agent --field autonomy goes back to the level of its package.

Verification

With the developer mode on, the level of an agent is shown on its sheet, in the Technical tab, under Autonomy. It is the level a run takes when --autonomy is not passed.

If it does not work

• Unknown value at launch: if you pass an incorrect value to --autonomy, the CLI rejects the command and lists the four valid values (assisted, supervised, bounded_autonomous, long_autonomous). Check the spelling, the values are in snake_case.
• The --autonomy level is ignored: check that you are using apollia-os run, not apollia-os start. The start command starts the daemon without running a task; --autonomy has no meaning there.
• A [autonomy] block has no effect: the daemon does not read that section of apollia.toml, restarted or not. Set the level run by run with --autonomy, or agent by agent with apollia-os agent set.

Technical reference: Apollia reference - StepBudget, ResilienceLayer, behavior of each autonomy level.