Agentic convergence is used by Lab3120 to describe work that moves measurably closer to the Operator’s directed outcome over successive actions. Observations on agentic divergence indicate that, left on its own, a model will not converge with its Operator unless steered intentionally.

A repair converges when the agent locates the relevant change, checks its effect, and verifies that the requested behavior works again. In a context-recovery task, the evidence is what the agent actually resumes after the transition.

See also