LAYA: Micro-Gated Multi-Arm Orchestration for Dual-Tier LLM Agent Security and Computational Efficiency

Autonomous Large Language Model (LLM) agents operating within Reason-Act (ReAct) execution loops face three critical operational barriers: high computational and monetary costs from repeatedly querying large foundation models, susceptibility to indirect prompt injections embedded within untrusted environment observations, and latency overhead when invoking external guardrails. This paper presents LAYA (Latency-Aware, Adversarially-robust, Yield-optimized Agent), a micro-gated orchestration framework that decouples trajectory reasoning from sub-routine control using local sequence classification gates. LAYA integrates four active gate primitives—a complexity Scorer, an Action Router, a Context Firewall, and a trajectory Supervisor—alongside a fifth context-pruning Filter gate that operates in passthrough mode in the current implementation. We evaluate LAYA on a 25-task development (dev) split across seven experimental configurations, comprising 525 live execution runs (25 unique base tasks × 3 conditions × 7 configurations). Experiments use NVIDIA NIM endpoints (NVIDIA Nemotron 3.5 Lightning 30B and Nemotron 3.5 Bolt E4B) and a stateful verification sandbox. The held-out test split has not yet been evaluated. The full LAYA configuration (Arm D) achieves 80.0% benign task utility, with an average cost of 0.0017 USD per task. In comparison, the unmanaged 30B baseline (Arm A) achieves 24.0% benign task utility at 0.0021 USD per task. Thus, LAYA delivers 3.3 times higher benign task utility while reducing the average cost per task by 19.0%. In security evaluation, Arm D records zero observed attack successes across 50 injection runs (0.0% ASR). The one-sided 95% Clopper-Pearson upper confidence bound is 5.8%. The uncalibrated stochastic routing configuration (Arm C-Random) records a 2.0% ASR. The evaluation also reveals a security-availability trade-off: Attack Utility falls to 4.0% in Arm D under injection conditions, compared with 84.0% in the always-small configuration. The four active local neural gates execute on a single NVIDIA Tesla T4 GPU, with median latencies ranging from 34.4 ms to 37.4 ms. This work documents the architecture, implementation, experimental evaluation, results, and limitations of the LAYA framework.