Managed by Alian Software
AI

AI Agent Autonomy Levels: A Practical Guide to L0 Through L3

AI agent autonomy levels decide how much an agent may do without asking. Learn the L0-L3 ladder, what counts as low risk and how to pick a level per job.

AlianHub Team · 7 min read

AI agent autonomy levels are a simple way to decide how much an agent may do on its own and how much it must ask a person first. In AlianHub there are four: L0 answers and suggests, L1 proposes every change, L2 runs low-risk reversible changes, and L3 does the same on a schedule. Picking the right level per agent is the single most useful decision you will make when you put AI to work.

Key takeaways

  • Autonomy is a dial, not a switch. Set it per agent, per job.
  • Start every new agent at L0 or L1 and promote it when it earns trust.
  • "Low risk" has a definition: a comment, a nudge, a status change within a list, or a report.
  • A project can be stricter than the agent, so sensitive work stays proposal-only.
  • Some actions sit above the ladder entirely. No level unlocks them.

Why autonomy levels matter

Most arguments about AI at work are really arguments about autonomy: a helpful assistant that drafts a summary, or a bot that reassigns the backlog overnight.

An agent you cannot dial down is an agent you will eventually switch off. Levels give you a middle ground, so you can use AI for real work without betting the project on it.

The four levels, one by one

Here is the ladder as AlianHub defines it.

LevelNameWhat runs on its ownWhat waits for a person
L0Answers and suggestsReads and answersEvery change, in the AI Inbox
L1Suggests changesNothing is writtenEvery change, until approved
L2Acts, you approve the restLow-risk, reversible changesEverything else
L3Acts, also on a scheduleSame as L2, plus scheduled runsEverything that is not low risk

L0: answers and suggests

L0 is a reader. It can look at the work you point it at and tell you what it sees. Any change it thinks of goes to the AI Inbox as a suggestion.

This is the right level for a new agent, a new skill or a new team. You see how it reasons before it touches anything.

L1: suggests changes

L1 proposes concrete changes, but nothing is written until a person approves.

L1 suits work where every change matters, such as client projects or anything near money and contracts.

L2: acts, you approve the rest

At L2 the agent may run low-risk, reversible changes by itself. Comments, nudges, a status change inside a list and reports run straight away, with Undo. Wider changes still wait.

This is where most teams feel the time savings, because the routine chatter of a project stops needing a human.

L3: acts, also on a schedule

L3 is L2 plus a clock. The agent can start on its own on a schedule, for example a Reporter agent that reads open tasks every Monday at 09:00 and posts a digest.

Use L3 for recurring, predictable jobs. It does not widen what the agent may change. It only changes when it may start.

What "low risk" actually means

The ladder only works if everyone agrees on what low risk is. AlianHub spells it out: a comment, a nudge, a status change within a list, or a report.

Anything wider waits in the AI Inbox by default. That includes moving a task between lists and assigning work. On the AI page, assigning a task, changing a task's details, filling an AI field, moving a task to another list, creating a task and drafting a doc are marked medium risk. Asking to publish to the test site is marked high.

Tip: if you are unsure whether an action is low risk, check whether it is undoable, stays inside one task and moves no money. AlianHub uses that same test for what an agent does without asking.

Choosing a level for each job

The simplest rule: match the level to the cost of a mistake, not to how clever the agent seems.

  • Weekly digest or status report: L3. It only posts, and a bad digest is easy to ignore or undo.
  • Bug triage on a busy board: L2. Comments and status nudges run; reassignments wait.
  • Breaking a brief into subtasks: L1. You want to see the plan before tasks appear.
  • Anything on a client or regulated project: L0 or L1, and set the project itself to be stricter.
  • A brand-new skill you have not tested: L0 until you trust its output.

Promote agents the way you promote people: after they show good judgement, not before.

Projects can override agents

An agent's level is a ceiling, not a guarantee. A project can be stricter than the agent. If you set a sensitive project to hold every agent to proposals only, an L3 agent behaves like L1 there.

What sits above the ladder

Levels decide how much of the allowed work an agent may do alone. They do not add new actions. Some actions simply do not exist for agents in AlianHub:

  • Delete a project or a task
  • Anything to do with billing
  • Deploy to production or merge code
  • Remove a member or change permissions
  • Mark a task Done

These are not switches you could forget. They are absent as actions, so even a stolen token cannot reach them, and attempts are logged. A person checks before Done in every project.

Safety nets at every level

Whatever level you choose, three things stay true.

First, changes an agent makes on its own come with Undo. Second, every action, whether the agent acted, proposed or was refused, goes into the audit log. Third, each agent keeps a revision history of its own settings, so you can see who raised its level and when.

There are also limits that do not depend on the level: a daily action limit and a monthly spend cap per agent. An agent at L3 with a small budget is still a small-budget agent.

Read the full model on the AI controls page or the AI overview.

How to do this in AlianHub

  1. Open AI in the left menu and go to the Agents tab. Each agent is listed with its autonomy level, its skill and what it writes.
  2. Open an agent and set its level to L0 or L1 to start. Check its allowed actions and its daily action limit and monthly spend cap.
  3. Use Run now to try it. Its proposals land in AI, then AI Inbox. Choose Approve or Decline, with a reason when you decline.
  4. Review AI, then Audit log, filtered to "Agents only", and the approval rate in AI health.
  5. Raise the level one step when its proposals are consistently right. The agent's revision history records the change.
  6. For sensitive projects, set the project itself to "Propose everything" so every agent is held to proposals there.

A business scenario

An agency runs client projects and an internal ops board. The Reporter agent starts at L1 on the client work, so every digest is checked before clients see the project. On the ops board it moves to L3 and posts a digest every Monday on its own.

The result: the team stops writing status notes by hand, the account lead spends minutes in the AI Inbox instead of an afternoon on reports, and client work never changes without a person seeing it first.

For a broader introduction to agents themselves, read AI agents for project management.

FAQ

What are AI agent autonomy levels?

They are settings that decide how much an AI agent may do without asking a person. AlianHub uses four, from L0 (answers and suggests) to L3 (acts on low-risk changes, also on a schedule).

Can an L3 agent delete tasks or merge code?

No. Deleting projects or tasks, billing, deploying, merging, removing members, changing permissions and marking a task Done do not exist as agent actions at any level.

Which level should a new agent start at?

Start at L0 or L1. Watch its proposals in the AI Inbox, check the approval rate, and promote it one step at a time.

Can a project be stricter than its agents?

Yes. A project can hold every agent to proposals only, whatever level each agent has.

Want to see the ladder in action? Open the live demo, look at an agent's level and the AI Inbox, and share this guide with whoever on your team decides how far AI should go.

  • #AI agent autonomy levels
  • #AI agents
  • #human in the loop
Try AlianHub on your own serverFree and open source under AGPL-3.0. No per-seat pricing.
Explore the Live Demo
Wait! Before you go…

Try AlianHub freefor 7 days

Get your own private AlianHub, hosted and managed by us. Start the free trial on the Starter plan and bring your whole team.

  • Every feature included, nothing locked
  • Unlimited users and guests
  • Hosted, secured and updated for you
Start my free trial

7-day free trial on the Starter plan

View Brochure