Adapting Architectures, Not Weights: Sequential Incumbent Replacement for Long-Lived Foundation-Model Agents

Foundation-model agents deployed in dynamic environments must adapt to persistent shift without repeatedly replacing working configurations on weak evidence. We study continual learning at the agent-system level: model weights stay fixed, while worker composition, communication, prompts, budgets, and verification change over time. We formalize this as sequential incumbent replacement with switching costs and hard constraints, and introduce Conservative Meta-Agent Search (CMAS), in which meta-agent refers to the entire outer loop and not to the proposal model alone. At each update, CMAS retrieves prior architectures and traces and proposes a bounded typed edit, then evaluates candidate against incumbent by paired shadow replay under common random numbers, assigns interventional credit, and promotes only when an empirically calibrated score clears a pre-specified margin, with a simulated canary stage and rollback. We evaluate on MetaInvest-Bench, whose controlled track contains recurrent, abrupt, and gradual regime shifts and a computable surrogate dynamic oracle. Across 96 streams, CMAS reduces normalized mean dynamic-oracle regret from $0.116$ to $0.104$, shortens post-shift adaptation delay from $3.1$ to $2.7$ updates relative to the strongest adaptive common-budget reimplementation, and raises held-out utility from $0.44$ to $0.48$. It retains 83.5% of utility on unseen regime mixtures and 81.2% with two new specialists; a backbone switch costs 7.8% utility versus 11.8% for AFlow-style search. False promotion is 3.4% per update and 4.8% in a sealed rerun, so update stability here is measured and not guaranteed. CMAS targets continual architecture adaptation and conservative update selection; weight-level catastrophic forgetting, backward transfer, and sequence-level error control are outside its scope.