LearnAIAgents
🏛️ Govern

Authority levels

Autonomy is earned, not granted.

Autonomy is earned, not granted

Authority is earned, not granted.You don't give a new hire the company credit card on Day 1.
LevelHuman roleAgent roleTrust earned by
1ObserveHuman actsAgent watches and learnsDay 1
2RecommendHuman decidesAgent suggests optionsWeek 1
3Act with ApprovalHuman approvesAgent executes approved actionsMonth 1
4Act and ReportHuman monitorsAgent acts autonomously, logs everythingQuarter 1
5AutonomousHuman auditsFull autonomy within defined scopeEarned

The delegation metaphor is the simplest way to remember this. You don't give a new hire the company credit card on day one. You give them a scope, you watch, you grant incrementally as they earn trust, and you demote them if things go wrong.

Agents are the same.

The five levels

  1. Observe. The agent reads but does not act. Useful as a shadow-mode launch before a more active deployment. Evidence accumulated here funds the next promotion.
  2. Recommend. The agent suggests options with reasoning; a human decides. Good for medium-risk, judgement-heavy tasks.
  3. Act with Approval. The agent executes pre-approved types of action on the human's say-so. Good for routine work with occasional irreversibility.
  4. Act and Report. The agent acts autonomously and logs everything. The human monitors in aggregate. Good for high-volume, low-individual-risk work.
  5. Autonomous. Full autonomy within a defined scope. Humans audit, not approve. Reserved for tasks where both the REMIT-E envelope and the REMIT-M monitoring are strong enough that the audit trail is the accountability.

Promotion requires evidence

An agent earns a promotion the same way a person does: evidence. What evidence?

  • Consistency. Success rate on the golden dataset above a threshold, for some rolling window.
  • Edge-case performance. Adversarial and ambiguous test cases not just passing but passing with the right reasoning.
  • Review-sample quality. Human-review disagreements at or below a threshold.
  • No regressions. Deploys have not regressed the metrics above since the last promotion.

None of this is novel. It is how you trust anyone in your organisation.

Demotion is always available

Behaviour can degrade. The model underneath can change. The world can change. Demotion back down the ladder must be as easy as promotion. Ideally, it is easier to go down than up.

A good governance process reviews the trust level of each agent at a fixed cadence (monthly for new agents, quarterly for mature ones), and has an emergency revocation path for when something unexpected happens.