# Agent definition evals

URL: /systems/trust-and-governance/controls/agent-definition-evals · Id: `control:agent-definition-evals`

> Every Change to an instruction, Tool, Skill, model Policy or model Version passes its Evaluator regression suite before merge.

Every [Change](/business/nouns/change) to an instruction, [Tool](/business/nouns/tool), [Skill](/business/nouns/skill), model [Policy](/business/nouns/policy) or model [Version](/business/nouns/version) passes its [Evaluator](/business/nouns/evaluator) regression suite before merge.

## Satisfies
| Criterion | Text |
| --- | --- |
| [SOC2 CC8.1](/systems/trust-and-governance/criteria/soc2/cc8-1) | The entity authorizes, designs, develops, configures, documents, tests, approves and implements changes to infrastructure, data, software and procedures. |
| [ISO27001 A.8.29](/systems/trust-and-governance/criteria/iso27001/a-8-29) | Security testing processes are defined and implemented in the development life of every system. |

## Implemented by
- [Verification](/code/verification)
- [Change verdict](/code/verification/change-verdict)
- [Evals](/agents/evals-system)

## Test
| Step | Rule |
| --- | --- |
| Kind | Preventive: it stops the failure before it happens. |
| Select | Every `Change.merged` Event whose target is an instruction, Tool, Skill, Policy or model Version. |
| Assert | Each cites a passing Evaluator verdict showing the new Version no worse than the one it replaces. |

## Owner
[Quality Lead](/agents/quality/quality-lead), [Quality Team](/agents/quality).

## Example
A Change rewrites the support [Agent](/business/nouns/agent)'s instructions for a recruiting [SaaS](/products/templates/saas). The Evaluator suite replays 400 past [Conversations](/business/nouns/conversation) against the new Version, which scores 0.2 points below the old one, so the Change cannot merge.