CAIN-42 CAIN Studio
CAIN-42 Academy

CAAE · Module 1 of 5 · 12 min

Agent CI: regression suites for behaviour

Agents regress silently: a prompt tweak, a model upgrade or a new tool changes behaviour. Agent CI runs fixed suites against a real agent endpoint on every change and diffs runs.

Make cases deterministic where possible: fixed inputs, pinned tool fixtures, recorded external responses. Score on outcomes and actions, not prose.

CAINBench runs regression suites and diffs any two runs.

Key points

  • Run suites on every change.
  • Pin inputs and tool fixtures.
  • Score outcomes and actions.
  • Diff runs.

In the CAIN-42 codebase

  • cainbench/ — regression suites and run diffs

Practice check

Practice 1 · choose one

Agent CI should run…

Practice 2 · true or false

“Scoring agents on the wording of their replies is the most reliable signal.”