Earned Autonomy: Every Agent Action by Level

The Earned Autonomy Ladder is the framework: five levels, moved by a record, with some actions kept for people. This page is how Nagent runs it. It lists every action an agent on the platform can take, the class that action belongs to, and the level at which an agent may take it without asking. It then sets out exactly what moves an agent up a level and what moves it down.

Every action belongs to one class, and a level decides what an agent does with each class.

An agent asks, through an approval card, for an action one class above its level. An action two or more classes above is not offered to it at all. A Locked action is always offered as a card and never done alone: deleting or taking something down, redirects, governance changes, lifting a check, spending past a cap, and creating agents, skills, workflows or triggers.

C5 Publish goes out alone at L5 only, and only through the checks publishing already has. Where a publishing path has no content check in front of it, its action stays Locked.

The list below is generated from the platform's action registry when this page is built, so it is the list the approval engine reads. Open an agent to see its actions.

546 actions across 98 agents, read from the platform's action registry when this page was built.

C1 Read, 72 actions

Reads. Changes nothing anyone keeps and spends nothing.

C2 Prepare, 215 actions

Writes only inside its own team's workspace: a draft, a plan, an analysis, a proposal, a report, or priced research inside its budget.

C3 Internal act, 159 actions

Changes workspace records other people rely on: a lead, a stage, a task, a setting, team memory, or a message to the team.

C4 External act, 54 actions

Reaches someone outside the workspace, or spends against caps on a live outside system: an email to a contact, a meeting invite, an ad budget.

C5 Publish, 18 actions

Puts words or pages in front of the public under the workspace's name, or changes the live site. Goes out by itself at L5 only, through its content checks.

Locked, 28 actions

Waits for a person at every level: deleting or taking down, redirects, governance changes, lifting a check, spending past a cap, and creating agents, skills, workflows or triggers.

Each agent arrives at the level its team was built for, shown on its page. From there, only a record and a person move it up.

1. The record builds up. Every action the agent asks for and a person approves, edits or dismisses is counted. So are the outcomes people rate and the thumbs they give.
2. The platform nominates. Each night it looks at every agent below L5. An agent qualifies after 30 days at its level with at least 25 decided cards, fewer than 5 per cent of them dismissed, and no automatic drop in those 30 days. It then files a "Promote to" card with that evidence.
3. A person decides, one level at a time. Raising a level needs permission to change AI governance. Raising to L4 or L5 also needs the autonomy edit permission.
4. L5 takes two people. A raise to L5 files a card that the person who asked cannot approve.
5. Every move is recorded. Each change writes one audit row: who, from which level, to which, why, and the evidence shown.

Nothing else raises a level. A rollback or a revert can only lower one, a re-provisioned team keeps the levels its agents had, and a long run of approvals on one kind of action is evidence on a nomination, never a promotion by itself.

A workspace can also open a training window for a new agent. For its first days or decisions, every action the agent would do alone waits for a person, whatever its level, and each decision teaches it.

The same nightly run lowers an agent whose record has slipped. It reads the last 14 days, never reaching back past the agent's last level change, and drops it one level if any of these hold:

• More than 30 per cent of its decided cards were dismissed or edited, with at least 10 decided.
• More than 15 per cent of its runs failed, with at least 20 runs. Failures caused by a provider outage are left out.
• Three different people gave its work a thumbs down.
• Two of its outcomes were rated poor.

It drops at most one level a night and never below L1. Three things drop an agent at once, without waiting for the night:

Every demotion writes the same audit row a promotion does, with what happened. Earning the level back takes a fresh record and a person's decision, the same way it was earned the first time.

An agent at a high level is not left unwatched. About 5 per cent of what each agent does alone, at most five actions a night, comes back to a person as a "Was this right?" card. The card does not hold anything up, and what the person says goes into the agent's record.

The level is a ceiling, not a promise. For any one call, an agent acts at the lowest of its own level, the workspace floor, a room's cap and a lead's own setting. A workspace can also mute an action, raise one action's bar, or put it on an approval policy. Each of these can only hold an action for a person. None of them can make an action run alone that its level does not allow.

• The Earned Autonomy Ladder: How much an AI agent may do alone, decided by its record
• State of Agent Governance: The control plane is a category before it is a product
• Anatomy of an Agent: A Configuration, Not a Program: The ten declared parts every Nagent agent is built from

• The Earned Autonomy Ladder, Nagent frameworks no. 1 https://nagent.ai/artefacts/earned-autonomy-ladder
• Nagent documentation, autonomy and governance https://nagent.ai/docs/autonomy-and-governance

Plain:

Nagent AI. Earned Autonomy: Every Agent Action by Level. Frameworks, no. 2. 2026. https://nagent.ai/artefacts/earned-autonomy

BibTeX:

@misc{nagent2026earnedautonomy,
  author = {Nagent AI},
  title = {Earned Autonomy: Every Agent Action by Level},
  series = {Frameworks},
  year = {2026},
  url = {https://nagent.ai/artefacts/earned-autonomy},
  note = {Published 2026-10-14, updated 2026-10-14}
}

The direct answer at the top of this page is written to be quoted as one sentence with this URL as its source.

Nagent is Multiplayer AI for end to end growth: a team of AI coworkers and your own people, working together in one workspace across marketing, sales and customer experience. Three things make it different. You approve the AI coworkers' work until they earn the right to act on their own. They carry the work all the way to pipeline and customers, not just content. And where your plan includes it, a Nagent marketer joins your team and owns the number with you. Founded in Bengaluru in 2024, Nagent is an Anthropic partner, holds four filed patents on orchestration and memory, and deploys in the customer's private cloud.

Published 14 October 2026. All rights reserved. Quote with attribution to Nagent AI and a link to this page.