# What evals decide

URL: /agents/how-agents-act/evals · Id: `page:agents/how-agents-act/evals`

> Evals grade every agent's work and gate every new version, and never change what waits for the founder.

Evals grade each [Agent](/business/nouns/agent)'s work and decide which version of it ships. What waits for the founder is set by [approvals](/agents/approvals), by action class and amount, and no eval score changes it.

## The record
Grades come from three sources: the [Role](/business/nouns/role)'s offline eval cases, [Evaluator](/business/nouns/evaluator) verdicts on live [Actions](/business/nouns/action), and the founder's corrections. Every edit, refusal or redirect of an Action becomes a new eval case. The [Evals](/agents/evals-system) record keeps every grade.

A falling pass rate opens a [Task](/business/nouns/task) for the seat's [Role Review](/business/nouns/role-review). The review ends in a [Decision](/business/nouns/decision): change the Role contract, tighten an approval, or seat another [Worker](/business/nouns/worker).

## 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](/business/nouns/startup)'s one [Metric](/business/nouns/metric), with guardrails.

A version that fails never ships. No Agent edits the Metrics, Evaluators or guardrails that grade it.

## One model for every Startup
Every Startup runs the same way: agents act, one-way decisions wait their delay windows, and watchers can stop them. A founder adds approvals by class and amount. Every Startup shares the same [Limits](/business/nouns/limit), spend states and legal floor.

## In practice
A [Services-as-Software](/products/templates/services-as-software) bookkeeping firm for dental practices approves `spend` above $100, so refunds under $100 run without the founder. In month three, one refund goes to a closed bank account and an Evaluator fails it. The case joins the suite, and the next Skill version must pass it before it ships.