Software Factories at enterprise scale: A federated platform for agentic developmentDownload the free Whitepaper

Invariants

What Are Invariants in AI-Assisted Change?

In short

Invariants are the properties that must not change, or did not change, when an agent makes a change. Stating them is cheap and saves a great deal of work: an agent that does not know what must hold will preserve the wrong things, and a reviewer who does not know what stayed the same has to check everything again.

Updated 24 September 20261 min read

Why do invariants help twice?#

They work in both directions of the handoff. Going out, they tell the agent what must hold, which is far more actionable than a list of things to do. Coming back, they tell the reviewer what they do not need to re-derive, which is what keeps a review proportionate to the change.

Keep up with the Knowledge BaseEvery two weeks, get new terms and updated definitions straight to your inbox.

Where do invariants belong?#

In the specification, alongside what the change should achieve — see spec-driven development. An invariant written only in someone's head is not available to the agent, and one written only in the review comment arrives too late to prevent anything.

Related terms

Found something wrong, missing or badly explained? Tell us. We credit every correction we publish.