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

Foundation-model agents deployed in changing environments must adapt from experience without repeatedly replacing working configurations on weak evidence. We study test-time continual adaptation at the agent-system level: after deployment, model weights remain fixed while the agent retains an experience archive and repeatedly updates worker composition, communication, prompts, budgets, model assignments and verification. We formalize this as sequential incumbent replacement with switching costs and hard constraints, and introduce Conservative Meta-Agent Search (CMAS). Throughout, meta-agent denotes the whole outer loop (proposer, compiler, evaluator, gate and rollback controller), not the proposal model alone. At each update, CMAS retrieves prior architectures and traces, proposes a bounded typed edit, 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. Update stability remains empirical: false promotion is $3.4$% per update and $4.8$% in a sealed rerun. Scope. CMAS targets architecture-level test-time continual adaptation and conservative update selection, not continual weight learning; we claim no internal skill acquisition, no prevention of weight-level catastrophic forgetting, no backward transfer and no sequence-level error control.