VERSIONING THE AGENT

Every Agent should have a version.

For example:

Agent: Research Agent

Version: 1.4

Model: Claude Sonnet 5

Instruction: v1.3

Knowledge: v2.1

Workflow: v1.2

When the output changes, you can know exactly which component changed.

The Agent improvement lifecycle

Don’t edit a running Agent directly and assume the new version will be better.

The process:

Version → Test → Compare → Evaluate → Human Review → Promote / Rollback

The goals:

• know which change truly improves quality;
• avoid fixing one error while creating another;
• keep a stable version to go back to;
• accumulate evidence before changing rules or methods.

This is a principle for designing and evaluating Agents.