A Two-Tier Perspective on Inference-Time Parallelism in Multi-Agent LLM Systems

Multi-agent systems (MAS) coordinate multiple LLM-powered agents through structured workflows, gaining reasoning power but incurring high inference latency from multi-step execution and repeated model invocations. Existing orchestration methods primarily optimize task performance and inference cost, leaving latency largely unaddressed. In MAS, end-to-end latency is governed by the critical execution path, so reducing total cost alone does not reliably reduce latency. Moreover, optimizing latency while preserving accuracy remains non-trivial: naive latency optimization can misassign operator-level credit and degrade task accuracy. To address this gap, we propose Latency-Aware Multi-agent System (LAMaS), a latency-aware orchestration framework for learning-based multi-agent systems. LAMaS addresses this challenge at two levels: at training time, it learns latency-aware execution graphs through constrained optimization with critical-path-aware credit assignment; at inference time, since a graph committed at training time cannot exploit runtime evidence, it complements graph construction with a lightweight controller that adaptively eliminates redundant future agent interactions as execution unfolds. Experiments on four benchmarks show that LAMaS achieves the best latency among evaluated learning-based MAS baselines, reducing end-to-end latency by over 50% while maintaining competitive or better accuracy. LAMaS is also modular and transfers to other MAS with minimal changes, consistently yielding latency reductions.

Multi-agent systems can solve complex tasks through collaboration between multiple Large Language Model agents. Existing collaboration frameworks typically operate in either a parallel or a sequential mode. In the parallel mode, agents respond independently to queries followed by aggregation of responses. In contrast, sequential systems allow agents to communicate via a directed topology and refine one another step by step. However, both modes are inadequate for achieving the desired objectives of minimizing communication and latency while simultaneously maximizing the accuracy of the final response. In this work, we introduce a hybrid paradigm called Nexa, a trainable response-conditioned policy that bridges the gap between the two modes. Nexa begins with a parallel execution stage, embeds the resulting responses into a shared semantic space, and then predicts a sparse directed acyclic communication graph. If the graph is empty, the system remains purely parallel; if it is non-empty, the system performs one sequential message propagation. The policy is a lightweight transformer model, and the method avoids the need for external LLM judges or reward models, as well as hand-crafted test-time topology search. We formalize this hybrid execution problem, show that the resulting graph is acyclic by construction, and that the framework strictly subsumes pure parallel execution, and present a training procedure based on policy-gradient optimization. Results demonstrate that the response-conditioned policy learned by Nexa under one setting can be reused when the number of agents, the task, or the underlying agent changes, thus emphasizing the generalizability of the learned communication policy.

Multi-agent tool calling is becoming the dominant interaction pattern for LLM-based systems, yet existing inference frameworks treat each tool call as an independent request, re-processing the entire conversation from scratch even though 85-95% of the prompt is unchanged from the previous turn. We present a stateful inference architecture that converts the $O(n_t)$ per-turn cost of conventional serving into an $O(Δ_t)$ delta-only cost: a persistent KV cache lives across turns and advances by ingesting only the new tokens, while a radix prefix cache extends this across interleaved multi-agent traffic and a prompt-lookup speculative decoder accelerates structured output. Against vLLM and SGLang on novel, fully-generated workloads, the reference implementation is $2.1\times$ faster per turn on a 6-turn agentic workflow and $4.2\times$ on the median turn of a 35-turn one, halving end-to-end wall time. The advantage comes from stateful reuse and speculation, not caching.