Skip to content
NagentNagent
Log inSign upHire your AI Growth Team
NagentNagent
Log inSign up
All documentation

Autonomy and governance

The five autonomy levels, the six classes of action, how a level rises and drops, the governance profile, what always asks a person first, and the per-agent budget caps that stop an agent when it reaches them.

Browse documentation

Autonomy is earned, not deployed

Every agent in your workspace works at a level. The level decides which of its actions run on their own and which wait for a person. An agent does not get a wider level because it was installed, or because time passed. It gets one because someone looked at its record and raised it.

Three things govern what an agent may do, and they stack:

  1. Its level, set per agent.
  2. The governance profile, set once for the workspace.
  3. Its budget, a daily and monthly cap on what it may spend.

Whichever is strictest wins.

The five levels

Nagent shows levels as L1 to L5, on Home, on Welcome and in the team rooms.

LevelNameWhat it means
L1ObserveReads and reports. Anything it would write, even a draft, waits for a person
L2AssistReads and drafts on its own. Changes to records other people rely on wait for a person
L3OperateChanges workspace records on its own. Anything that reaches outside the workspace waits for a person
L4RepresentActs outside the workspace on its own, such as emailing a contact. Publishing waits for a person
L5LeadPublishes on its own and can hand work to agents below it. Deleting and governance changes always wait for a person

The Agent Workbench, counting the workspace's agents at each level from L1 Observe to L5 Lead, with a filter per level

Every action has a class

Each action an agent can take belongs to one of six classes. A level runs every class up to its own number on its own, and holds anything above it for a person.

ClassWhat puts an action in itRuns on its own from
C1 ReadReads. Changes nothing anyone keeps and spends nothingL1
C2 PrepareWrites only inside its own team's workspace: a draft, a plan, an analysis, a reportL2
C3 Internal actChanges workspace records other people rely on: a lead, a stage, a task, a setting, team memoryL3
C4 External actReaches someone outside the workspace, such as an email to a contact or a meeting inviteL4
C5 PublishPuts words or pages in front of the public under your name, or changes the live siteL5
LockedDeleting or taking down, redirects, governance changes, lifting a check, spending past a cap, and creating agents, skills, workflows or triggersNever: always waits for a person

On an agent's Advanced tab, the actions table shows each action's Auto floor: the level that runs it on its own and the class it belongs to, or Locked. Every action Nagent ships, listed by class, is on the public page Earned Autonomy. See approvals and the Inbox for what happens to an action that waits.

The level your teams start at

A team you install starts its agents at L4 Represent, so it can work from its first message rather than asking about everything. Know what that does not mean for the Sales team: sending e-mail is gated separately from the level on NORA's seat. On a new workspace every outbound send waits for a person's approval, whatever level her seat shows.

Changing an agent's level

Open the agent in the Agent Workbench. Its header carries two buttons for this, beside Stop agent:

  • Promotion review moves the agent one level up. It shows the agent's evidence for the last 30 days, and what changes, today and after. Write why the agent is ready, which is kept on its permanent history, and press Promote to the next level, or Hold for now.
  • Change level moves the agent one level down. Write why, and press Move to the level below, or Cancel.

Promotion review for an agent at L4 Represent: its evidence for the last 30 days, what changes from L4 Represent to L5 Lead, the box for why it is ready, and Promote to L5 LEAD or Hold for now

The same levels can be set from the Configuration tab: on its Instructions page, under Governance, with the Autonomy level control and Save governance, and on its Autonomy page, under Set new level. Every screen names the levels the same way: L1 Observe, L2 Assist, L3 Operate, L4 Represent and L5 Lead.

An agent's page in the Agent Workbench: Pause and Sandbox, its status, earned autonomy, results and risk, and Promotion review, Change level and Stop agent below

The rules for a move are the same wherever you make it:

  • Down can skip. An agent can be lowered any number of levels at once, by anyone allowed to change agents.
  • Up is one step at a time. Asking for more is refused: "A level goes up one step at a time: from L2 the next is L3."
  • L4 and L5 need a lead's role. Raising an agent to L4 or L5 is open to an Organisation admin, Chief of staff or Team lead, and not to a Member.
  • L5 follows your workspace's approver setting. With one approver, the default on every plan, the raise to L5 applies at once. With two approvers switched on, a raise to L5 files an approval card, "Raise" the agent "to L5?", and someone other than the person who asked approves it. See two approvers.

Every move is recorded with who made it and why. Agents the platform maintains for every workspace, such as the builders, can be switched on or off and paused, but their level and set-up are fixed. If your role can read a level but not change it, the agent's page says so.

Once a workspace has run for a quarter, Home counts the recorded moves under Autonomy level changes, 90 days, as rises and drops, with a link to Open the autonomy ledger. Each move is also in the audit log.

Two approvers

Two approvers is a setting for each workspace. It is off by default on every plan, Enterprise included, and a platform admin at Nagent switches it on for a client who asks for it. Your own admins cannot switch it on or off.

  • Off, one approver. Anyone who can approve held work approves it alone, their own requests included, and a raise to L5 applies at once.
  • On, two approvers. The person who asked for something never approves their own request; someone else who can approve does. A raise to L5 files an approval card for someone else, and so does approving a nomination to L5. If you try to approve your own request, Nagent says someone else approves it in this workspace.

If you are the only person who can approve and two approvers is on, your own L5 raise waits with this note: "L5 needs a second person here: the one who asked cannot approve it. A platform admin can switch two approvers off." Either invite a second person with a role that may raise an agent to L5 (Organisation admin, Chief of staff or Team lead) to approve it, or ask Nagent to switch two approvers off.

Who may approve at all does not change: that is still your role, and on MOXA the approvers you name.

How a level rises

A promotion is nominated and decided, never automatic. Each night Nagent reads every agent's record. An agent that has been at its level for 30 days, has had at least 25 of its approval cards decided, has had fewer than 5 in every 100 dismissed, and has not been dropped in that time, is nominated: an approval card, "Promote" the agent "to" the next level, goes to the Organisation admins. A person decides, and the same permission rules apply as above.

The same nightly run sends a small sample of what agents did on their own back to a person as Was this right? cards. They block nothing: approving one says the action was right, and dismissing it says it was not, which counts against the agent's record.

How a level drops

Each night, an agent drops one level when its last 14 days show any of these:

  • More than 30 in every 100 of its decided cards dismissed or edited, with at least 10 decided.
  • More than 15 in every 100 of its runs failing through its own error, with at least 20 runs. Errors caused by a provider outage are left out.
  • A thumbs down from three or more different people.
  • Two or more outcomes rated poor.

It drops at most one level a night, and never below L1. Some drops are immediate:

  • A budget breach drops the agent one level.
  • An attempt at a Locked action outside an approval card drops it to L1 and pauses it.
  • Something an L5 agent published being taken down drops it to L4.

The autonomy ledger

Governance is the autonomy ledger: how far each agent has earned its autonomy. It counts the agents scored and how many are healthy, on watch, and at risk or critical, and lists every agent with its trust, karma, errors, blocks and autonomy, filtered by band. Each row opens the agent's page. Nothing on the ledger changes a level; it is where you look before you do.

The autonomy ledger: the governance profile, predictive warnings, cognitive drift and operator impact across the fleet, and every agent's score, lowest first, with its trust, karma, errors, blocks and autonomy

The governance profile

The profile sets the workspace's floor for the website and content actions the visibility agents take, on top of each agent's level. There are two.

ProfileWhat it says
StandardAgents act within their autonomy level and ask first on anything destructive
BFSIStricter defaults for regulated work; more actions wait for a person

Until you choose, Standard applies, and Workspace set-up shows the Autonomy and governance card as Needs you, because nobody has said so.

To choose, open Governance. The Governance profile control sits at the top of the page. Press the profile you want.

How the profiles treat each kind of website and content action:

Kind of actionStandardBFSI
ReadsRunRun
Metadata fixes, internal linksRun once the agent is at L2Run once the agent is at L3
Creating or editing contentAlways asks, below the Enterprise tierAlways asks
Changes to priority settingsAlways asksAlways asks
Destructive changesAlways asksAlways asks

What always asks first

Some actions ask a person every time, at every level, under either profile, and no promotion can change that.

  • Destructive changes: adding a canonical tag or a noindex directive, the changes that can take a page out of search or point it elsewhere.
  • Changes to how work is prioritised in the workspace.

An agent with no governance record at all also queues every action with a side effect, rather than acting unchecked.

And at any level, an agent that is paused or switched off takes no action with a side effect. Pausing is the control to reach for when something looks wrong: it takes effect on the next action, and nothing the agent was approved for earlier overrides it.

Our logic

Some mistakes can be undone and some cannot. A meta description written badly can be rewritten tomorrow. A page removed from search may not come back. The first kind is where an agent can earn its way to working alone. The second kind always has a person on it.

Budget caps

Every agent has a spending cap, a daily one and a monthly one. Until you set your own, each agent carries the default: 1,000 credits a day and 20,000 credits a month.

Before an agent makes a paid call, Nagent checks what it has left under both caps, counting money already committed to calls still running. If the call would go over, it is refused.

Setting a cap

  1. Open the agent in the Agent Workbench and go to its Budget pane.
  2. Read Today and This month: spend against each cap.
  3. Under Edit caps, set Daily cap (credits) and Monthly cap (credits).
  4. Choose When exceeded, then press Save.

If the agent has no budget of its own yet, the pane says so and pre-fills the form. Check When exceeded before you save: the form pre-fills Pause the agent, while an agent with no saved budget behaves as Queue actions for approval.

What happens when a cap is reached

When exceededWhat happens
Queue actions for approvalFurther paid calls are refused and the attempt is recorded as held. The agent starts again when the day or month rolls over
Pause the agentThe agent is switched off and stays off until a person turns it back on
Notify only (continue)The agent keeps running and the breach is announced

The pane's status reads Healthy, Approaching cap, Exceeded or Paused. Daily caps reset each day and monthly caps each month.

Whichever you choose, reaching a cap also drops the agent one autonomy level, as described under how a level drops.

The workspace total on Home and on the set-up Budget and spend card is the sum of these per-agent monthly caps. It is what the workspace can spend before its agents stop, not a plan allowance.

Replies agents give in a team room conversation are counted today under a shared system entry rather than against the replying agent's own cap. They are still recorded and still priced.

Our logic

A budget is a stop, not a report. A figure you read at the end of the month cannot prevent anything, so the check happens before the call, and the default behaviour is to hold the work rather than to keep spending.

Where to go next