# Agent autonomy should be earned per client

> ARC uses four levels of agent autonomy: shadow, suggest, act within policy and act with sampling. Recovery operators can choose a level for each agent and client, review the evidence and move back to approvals whenever results or rules change.

Canonical: https://arc-repo.com/blog/agent-autonomy-ladder

---

At ARC we review autonomy per agent and client. Our practice is to start in shadow, approve each action and promote only when the evidence supports it. That gives every level a clear way to be reduced again.

The levels are shadow, suggest, act within policy and act with sampling. They give the team a practical way to expand automation while keeping authority explicit.

## L0 shows the work without writing it

Shadow mode gives the team a useful answer before it gives the agent authority. The agent evaluates assignments and shows what it would have done, while writing nothing to the operator's systems.

That makes shadow a practical starting point for a representative week of assignments. The review asks which actions would have happened, where the agent would have stopped and which decisions would have come to a person.

It also exposes missing context. If the agent can't see a current rule, a required source or an independent hold, the team learns that before enabling a mutation.

## L1 keeps every action behind approval

Suggest mode lets the agent prepare the next action while a person approves each one. The reviewer sees the evidence attached to the proposal and records the decision under their own name.

This is where the team learns whether the agent's recommendations are useful in real cases. It is also where exceptions become visible: the cases that need judgment should arrive with a clear reason and next owner, and stay out of the automation log.

## L2 acts inside delegated policy

At the next level, the agent may act within configured policy and route warnings to a person. Delegated authority still comes from current provisioned grants and client-specific rules. A model instruction doesn't create permission.

Every mutation uses the same governed action surface. The record exposes the assignment, trigger, reason and outcome, so autonomy doesn't make the action less accountable.

## L3 keeps score and stays reversible

Sampled autonomy lets the agent act while its work is reviewed at the configured rate. The team watches the results and decides whether that level remains justified.

Promotion isn't permanent. A client change, policy change or weak result can move the agent back to approvals or shadow. Independent stops don't move down the ladder with the agent; they retain their own fences and acknowledgment obligations.

L0 is shadow, L1 approves each action, L2 acts within policy and L3 acts with sampling. The whole ladder stays available when a client needs a different level.

Figure: Autonomy increases only when the client and evidence support it.

```mermaid
%% caption: Autonomy increases only when the client and evidence support it.
flowchart LR
  A[L0 Shadow] --> B[L1 Approve each action]
  B --> C[L2 Act within policy]
  C --> D[L3 Act with sampling]
  D -->|Numbers weaken| C
  C -->|Policy or client changes| B
  B -->|Need a clean comparison| A
  class A mark
  class C accent
  class D mark
```

This gives the recovery operator a measured way to expand automation without turning a successful pilot into an unreviewed default, because each client can start with evidence, promote only the work that performs well and keep every approval and restriction attached to its decision. Autonomy always has a way back.

Consider an illustrative week of intake work. On Monday the agent reads ARC-1042 and proposes the same intake action a reviewer would have taken, but writes nothing. On Tuesday the reviewer promotes only that action to approvals. By Friday the team can compare the proposed action, the approved action and the exception that still needed a person. The client has evidence about one specific step, not a vague story about automation. L0 and L1 did their jobs before L2 ever wrote a record.

That progression is the point.

## Frequently asked questions

### Does every client have to use the same autonomy level?

No. Autonomy is set per agent and per client. One client may allow policy-bound intake while another requires approval for the same action.

### Can a model prompt grant more authority?

No. Instructions can guide work, but they don't grant authority. Current identity, provisioned grants, policy, required approvals and independent stops govern what the agent may do.

### What if the agent's results get worse?

The team can lower autonomy at any time. L3 is conditional on the evidence supporting it, not a permanent destination.

See how ARC governs agent actions or bring a week of assignments for shadow mode.

Read this page as Markdown