Autonomy should be earned: the five Levels of Agency
An agent should not go from a demo to unrestricted authority overnight. Here is how we let it earn more, one level at a time, and what has to be true before each step up.
Most agent projects fail in one of two ways. Either the agent never leaves the pilot, because nobody can say when it is safe to let it act, or it is given broad access on the first day and the first bad action ends the project. Both come from the same missing idea: that authority is something an agent earns, step by step, with evidence.
We use five Levels of Agency for that. Every agent we deploy holds exactly one level in each workflow, and it moves up only when the record supports it.
The five levels
Observe
The agent watches the workflow and builds an understanding of it. It has read access to the systems the work touches and no ability to change anything. People do all of the work, and what the agent sees goes to the team designing the system.
Recommend
The agent proposes decisions and actions, and people remain responsible for carrying them out. Every recommendation carries its evidence and the result of the policy check, so the person accepting or overriding it can see why it was made.
Human-approved execution
The agent prepares actions and carries them out only after a person explicitly approves. The approver sees the evidence, the policy result and the agent's confidence before deciding, and every approval and override is written to the record.
Confidence-gated autonomy
High-confidence actions that comply with policy run automatically. Anything uncertain or exceptional escalates to a person with the evidence attached. People still review a sample of what ran on its own.
Governed autonomy
The agent operates approved parts of the workflow on its own, inside the permissions, policies and controls set for it. People own the workflow, set the limits, review performance and take every case the agent escalates.
What moves an agent up
A level is not a reward for time served. Before an agent starts, we agree with the workflow owner what has to be true for each step up. Typically that means recommendations accepted without changes for the cases in scope, confidence that lines up with correct outcomes, exceptions that are understood, and recovery that has been tested.
Then we review the evidence together. Authority can go down as well as up: if performance slips outside the limits the owner set, the level is reduced.
Why the level is per workflow
The same agent can hold different levels in different workflows. Reading invoices and proposing adjustments might run with confidence-gated autonomy while releasing a payment never leaves human hands. The level describes what the agent may do, and that depends on the risk of the work, not on how clever the model is.
Where to start
Most workflows start at Observe or Recommend. That is not caution for its own sake. It is how the record that justifies the next level gets built. If you want to know which of your workflows could start higher, that is what our audit is for.