# Autonomy earned by evals

URL: /agents/autonomy-levels/evals · Id: `page:agents/autonomy-levels/evals`

> In a Startup with a human founder, each agent earns authority one decision class at a time, from assist to supervised to full.

In a [Startup](/business/nouns/startup) with a human founder, an [Agent](/business/nouns/agent) earns authority one decision class at a time. It climbs from `assist` to `supervised` to `full` on its eval record and drops back automatically when the record slips.

## Three levels
| Level | The Agent | The founder |
| --- | --- | --- |
| `assist` | Drafts each [Action](/business/nouns/action). | Sends, edits or discards the draft. |
| `supervised` | Acts. Every Action waits a delay window in the founder's pending feed, even a two-way one. | Reads the feed, and may pause, redirect or cancel. |
| `full` | Acts under the class's normal window. | Reads only what they choose. |

A [Role](/business/nouns/role) holds one level per decision class. The [Success Manager](/agents/customer/success-manager) can be `full` on support replies, `supervised` on refunds and `assist` on price changes, all at once.

## The record that promotes
Grades come from three sources: the Role's offline eval cases, [Evaluator](/business/nouns/evaluator) verdicts on live Actions, and the founder's corrections. At `assist` and `supervised`, every edit, redirect or cancel counts as a failure and becomes a new eval case. The [Evals](/agents/evals-system) record keeps every grade.

A class moves up one level when the Role passes 95 percent or more of its last 50 graded Actions there. It moves down one level the moment that rate falls below 95 percent. A breach of a [Guard](/business/nouns/guard) or the legal floor drops it straight to `assist`, and every move is a [Decision](/business/nouns/decision) on the record.

## New versions earn their way in
[Skills](/business/nouns/skill), prompts, Role policies and [Mandates](/business/nouns/mandate) are versioned definitions with their own eval suites. A new version must score no worse than the current one offline before it touches a customer. It then runs as an [Experiment](/business/nouns/experiment) against the Startup's one [Metric](/business/nouns/metric), with guardrails.

A version that passes keeps the levels its Role has earned. A version that fails never ships. No Agent edits the Metrics, Evaluators or guardrails that grade it.

## Mixing the two paradigms
A zero-human Startup skips the ladder: every class runs at `full` and relies on delay windows and watchers alone. A [Single human](/agents/operating-models/single-human) Startup mixes the two. Its agents run `full` where they have earned it, and the founder watches the rest.

Every Startup shares the same [Limits](/business/nouns/limit), spend states and legal floor. Autonomy changes who watches; it never removes a gate.

## In practice
A [Services-as-Software](/products/templates/services-as-software) bookkeeping firm for dental practices starts its [Billing Specialist](/agents/finance/billing-specialist) at `assist` on refunds. The founder edits 6 of the first 40 drafts, each one a new eval case. Once 48 of its last 50 refunds pass, the class moves to `supervised`, and five weeks later to `full`.

In month three, one refund goes to a closed bank account and an Evaluator fails it. The rate dips to 94 percent, so refunds return to `supervised` that hour. The founder watches 50 more, the rate recovers, and the class climbs back without a meeting.