Skip to content
ML Academy

Introduced on day 49

Stopping condition

A rule enforced by the harness that ends an agent run, such as a verified done state, an exhausted budget, a repeated call, consecutive tool errors or a need for human input.

In detail

Each condition returns a distinct reason code so that runs can be analysed by how they ended. A done signal from the model should be accepted only if a check of the expected state passes. Any stop other than a verified done returns a partial result and is never presented as complete.