Conservative Meta-Agent Search: Reversible Architecture Updates under Non-Stationarity

Persistent AI agents operating in changing environments must adapt their workflows without repeatedly replacing functioning configurations on weak evidence or violating fixed operational constraints. We study continual adaptation at the agent-system level: foundation-model weights remain fixed while worker composition, communication, prompts, model assignments, budgets, aggregation, verification, and an archive of prior experience evolve. We formalize this as sequential incumbent replacement with switching costs and hard eligibility constraints, and introduce Conservative Meta-Agent Search (CMAS), in which meta-agent denotes the whole outer loop (proposer, compiler, evaluator, gate and rollback controller) rather than the proposal model alone. At each update, CMAS retrieves prior architectures and traces, proposes a bounded typed edit, evaluates candidate and 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; promotion is followed by a simulated canary stage and rollback. Evaluation uses MetaInvest-Bench, a risk-sensitive asset-management testbed combining controlled non-stationary streams, leakage-controlled market replay, and a governance suite. Across 96 controlled streams, and relative to the strongest adaptive common-budget reimplementation, 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, and raises held-out utility from $0.44$ to $0.48$. False promotion is 3.4% per update and 4.8% in a sealed rerun. Across 576 simulated governance incidents, CMAS recovers within three updates in 78% of episodes, with rollback precision/recall $0.85/0.80$. Deployment inference cost is lower than the strongest adaptive reimplementation, but CMAS is heaviest in search-inclusive tokens and we claim no net compute saving. The scope is controlled, reversible continual adaptation of persistent agent systems; continual weight learning, live enterprise deployment, weight-level catastrophic forgetting, backward transfer, and sequence-level safety guarantees all lie outside it.