Position: Adapt Fast, Evolve Slowly: Evidence-Gated Escalation for Agentic Continual Learning

Agentic continual learning increasingly allows and encourages systems to revise not only model behavior but also memory, routing, workflows, harnesses, and even their own adaptation mechanisms. Yet greater persistence is not inherently better: a transient failure may be solved through local recovery, while arbitrary in-depth modification can expand the consequences of an incorrect update. We take the position that agentic continual learning should follow an adaptation escalation principle that uses the least persistent intervention sufficient for the available evidence, and escalate, by default, only when shallower adaptation proves inadequate. The burden of proof should lie with deeper and more costly adaptations: continual learning should escalate because the evidence demands it, not merely because the system can. We organize this principle by looking at modifications at three timescales: within-episode runtime adaptation, cross-episode heuristic learning, and structural self-evolution. We argue that the final gate of escalation decisions should depend on independently assessed consequences of the change, and further propose an architecture for implementing gated escalation in agent systems. Finally, we make the position falsifiable through matched lifecycle baselines, protected and independent evaluation, escalation stress tests, and reproducible update lineages. This framing shifts continual agent improvement from maximizing adaptation toward justifying when deeper adaptation is warranted.