Continual Learning for Enterprise Agents is a Change-Control Problem

Enterprise AI agents operate where models, prompts, knowledge bases, schemas, tools, permissions, and evaluation harnesses evolve asynchronously. We argue that enterprise continual learning is a governed state-transition claim rather than a property inferred from component change alone. We formalize agent configuration as a version-addressed six-part state spanning model, knowledge, ontology, tools, policy, and evaluation context. An adaptation candidate pairs an explicitly scoped intended change set with an observed manifest difference. The two must be reconciled through two-sided checks; the candidate is then evaluated under a predeclared contract using custody-separated evidence and checked for cross-component effects, after which an authorized role decides whether to promote it. We illustrate this framing with a bounded knowledge-acquisition case: a development-selected candidate achieved higher development-set micro-F1 (0.4444 vs. 0.2105), but recovered none of eight target mentions under frozen zero-tuning transfer (precision, recall, and micro-F1 0.0000). Lacking a same-split transfer baseline, this outcome does not demonstrate negative transfer; it shows only that the development gain did not support contextual generalization. We outline research directions in state capture, attribution, benchmark custody, authority drift, and reversible adaptation.