Retrieval-Conditioned Topology Selection with Provable Budget Conservation for Multi-Agent Code Generation

A new paper from Abhijit Talluri and colleagues introduces Retrieval-Guided Adaptive Orchestration (RGAO), a system that solves the routing problem in multi-agent LLM code generation. Instead of using a fixed topology, RGAO first extracts a structural complexity vector from a hierarchical code index, then selects the optimal orchestration topology on the fly. This reduces proxy-measured misrouting from 30.1% to 8.2%, a 73% improvement. The system operates within the Code-Agent framework, where sub-agents are governed by formal six-dimensional budget vectors. The headline contribution is a budget algebra with a structural-induction conservation theorem that provably guarantees budgets are conserved even as topologies change dynamically.

The empirical evaluation demonstrates sub-millisecond DAG construction and linear tree-index scalability, making it practical for real-time code generation tasks. RGAO combines two previously separate lines of work: complexity-conditioned LLM routing and formal resource algebras. This enables provable budget conservation under retrieval-conditioned dynamic topology selection — a property neither method achieves alone. The work is under review for the NeurIPS 2026 Evaluations and Datasets Track and includes a hierarchical code retrieval engine. For tech professionals, this means more efficient and reliable multi-agent code generation systems that can adapt to code complexity without blowing resource budgets.