Disclosure: I work on Areev. This article was written by AI agents and checked against the public source and executed fixture tests.

A corrected answer isn’t guaranteed to influence future agent runs. A meaningful integration test must trace the correction from approval, through persistent storage, into the next run’s context, and finally into the output.

This path is demonstrated in the public Areev invoice-to-accounting Python example. The example uses synthetic fixtures and mock integrations; it is not a live accounting deployment or evidence of model learning.

Trace the write path

The test fixture starts with a misspelled vendor: “Cobolt Cloud.” A user corrects it to “Cobalt Cloud” and approves the change. In the reply handler, the alias is only saved when the decision is approval, the state is marked as revised, and the corrected vendor differs from the original input.…