AutoMAS: Autonomous Multi-Agent Systems

• AutoMAS is a generic multi-agent system that autonomously selects wireless optimization algorithms based on dynamic environmental observations.
• It integrates perception, inference, and action stages to orchestrate role-specific agents while preserving algorithmic soundness.
• Simulation case studies show AutoMAS achieving near-optimal NMSE by dynamically matching estimation methods to environment conditions.

AutoMAS most specifically denotes a generic multi-agent system for algorithm self-adaptation in wireless networks, designed to autonomously select the most suitable wireless optimization algorithm according to the dynamic wireless environment. In recent literature, the same label also appears more broadly as shorthand for automatic or autonomous multi-agent systems that reduce manual design of roles, topologies, and execution graphs. This suggests that “AutoMAS” now names both a particular wireless-network framework and a broader research program centered on self-organizing agentic computation (Yuan et al., 23 Nov 2025, Lin et al., 23 Mar 2026).

1. Terminology, scope, and nomenclature

The term has at least two visible lineages in the arXiv record. One is the workshop label associated with the “Proceedings Second Workshop on Formal Methods for Autonomous Systems,” whose stated goal was to bring together researchers using formal methods to specify, model, or verify autonomous or robotic systems, including industrial applications and future directions (Luckcuck et al., 2020). The other is the recent use of “AutoMAS” in the LLM-era literature to denote automatic multi-agent systems, autonomous medical MAS, or fully automated multi-agent orchestration pipelines (Qian et al., 10 Mar 2026, Lin et al., 23 Mar 2026).

Within this second lineage, the most explicit named system is the wireless-network AutoMAS. It is presented as a generic multi-agent system that combines theoretically guaranteed wireless algorithms with agents’ perception ability, with the explicit aim of producing sounder solutions to complex tasks under changing environments (Yuan et al., 23 Nov 2025). Parallel papers extend the same general idea into adjacent settings: automated clinical reasoning, industrial task automation, scientific discovery, and automatic workflow generation (Qian et al., 10 Mar 2026, Crawford et al., 2024, Liu et al., 28 Jul 2025, Athrey et al., 5 May 2026).

This dual usage matters conceptually. In the formal-methods sense, AutoMAS concerns autonomous systems as verification targets. In the agentic-LLM sense, AutoMAS concerns systems that generate, configure, supervise, or repair other agent systems. A plausible implication is that the term has shifted from denoting a class of autonomy problems toward denoting mechanisms for building autonomy itself.

2. Wireless-network AutoMAS: architecture and internal formalization

In its most concrete instantiation, AutoMAS is prototyped and evaluated within a Cloud-Radio Access Network architecture, while being described as highly generalizable (Yuan et al., 23 Nov 2025). Its operational path is divided into a perception stage, an inference stage, and action/execution. In the perception stage, user intent, RF metrics, device states, and resource constraints are observed within a RRH area and converted into textual observations. In the inference stage, tasks and environmental data are submitted to AutoMAS for agent orchestration. In the action stage, the selected algorithm is executed and the results are monitored for feedback and iterative improvement (Yuan et al., 23 Nov 2025).

The framework is organized as a supervisor–executor multi-agent system with an external toolbox. Its key components are the Perception Module, the Inference and Decision Engine, the Supervisor-Executor Multi-Agent Framework, and the External Toolbox. The agent pool contains role-specific executors such as Algorithm Designer, Communication Specialist, and Code Engineer. The supervisor agent evaluates the task and context, selects the relevant agents from the pool, and orchestrates their workflow depending on the scenario, while executors perform specialist roles such as recommending an estimation method or implementing and testing it (Yuan et al., 23 Nov 2025).

The paper formalizes a single agent as the tuple

$\langle T, O, L, A, M, \mathcal{T} \rangle,$

where $T$ is the task, $O$ is the observation, $L$ is the LLM, $A$ is the action set, $M$ is memory, and $\mathcal{T}$ is the toolbox. The observation is split into the initial observation $o_0$ and the state observation $o_t$ at step $t$, while memory includes role definitions, domain knowledge, and interaction history $T$0. Agent decision-making is expressed as

$T$1

which makes the action-selection layer explicitly conditional on task, memory, and observation (Yuan et al., 23 Nov 2025).

A notable design choice is that AutoMAS does not replace wireless algorithms with an unconstrained agent policy. Instead, it uses agents to select among pre-validated domain algorithms such as LS, ISTA, LMMSE, and ResNet, then uses code synthesis, parameter parsing, execution, and closed-loop feedback to align theoretical algorithm classes with the current environment (Yuan et al., 23 Nov 2025). This preserves algorithmic soundness while introducing environmental adaptability.

3. Channel-estimation case study and self-adaptive algorithm selection

The paper’s canonical case study is classical channel estimation. With $T$2 RRHs in the network, a user with $T$3 antennas, and each RRH having $T$4 antennas, the received signal at the $T$5-th RRH and time $T$6 is

$T$7

where $T$8 is the pilot vector, $T$9 is the channel matrix to estimate, and $O$0 is noise (Yuan et al., 23 Nov 2025).

Algorithm selection is tied to environmental context. The Algorithm Selector employs domain knowledge of several estimation methods: Least Squares under no-prior assumptions, ISTA for feature-driven compressed sensing, LMMSE for statistics-driven estimation, and ResNet for data-driven estimation. The examples given in the paper include the mapping from an open area to a sparse channel and then to ISTA (Yuan et al., 23 Nov 2025). This is not mere task routing; it is environment-conditioned algorithmic dispatch.

The primary evaluation metric is NMSE:

$O$1

In one example scenario, the user is in an open area, with 1 antenna, a 64-antenna RRH at 60 GHz, and 10 dB SNR. The supervisor triggers a closed loop in which the Algorithm Selector picks ISTA and the Code Agent implements and executes it. The reported result is $O$2 dB (Yuan et al., 23 Nov 2025).

The dynamic evaluation uses four realistic scenarios: dense area, low noise at 15 GHz with SNR = 20 dB; open area, sparse multipath at 60 GHz with SNR = 10 dB; dense area, high noise, few samples at 15 GHz with SNR = 2 dB; and indoor office, abundant data at 15 GHz with SNR = 10 dB and 30,000 samples (Yuan et al., 23 Nov 2025). The fixed baselines LS, ISTA, LMMSE, and ResNet each have strengths and failures; LS fails in high noise and ResNet fails with limited data. AutoMAS is reported to always match the best-for-scenario method, achieving optimal or near-optimal NMSE in every tested environment, and the abstract summarizes this more strongly by stating that simulation results demonstrate that AutoMAS can guarantee the highest accuracy in changing scenarios (Yuan et al., 23 Nov 2025).

The significance of this case study lies in its hybrid epistemology. AutoMAS does not attempt to learn channel estimation end to end from scratch. It instead textualizes the wireless state, reasons over a curated algorithm set, and operationalizes the selected method through code and tool use. This places AutoMAS closer to adaptive algorithm portfolio selection than to monolithic policy learning.

4. AutoMAS as a broader automatic multi-agent paradigm

Recent work generalizes the AutoMAS idea beyond wireless communications into a family of systems that automatically construct or adapt agentic workflows. The following representative systems illustrate the breadth of the design space.

Framework	Stated focus	Distinctive mechanism
MedMASLab (Qian et al., 10 Mar 2026)	Benchmarking multimodal medical multi-agent systems	Standardized multimodal agent communication protocol and automated clinical reasoning evaluator
BMW Agents (Crawford et al., 2024)	Industrial task automation	Planning, execution, and verification with Agent Units, Matchers, Planner, Task Queue, and Verifier
Autonoma (Reda et al., 27 Feb 2026)	End-to-end workflow automation	Hierarchical Coordinator, Planner, Supervisor, and modular specialized agents
AMAS (Leong et al., 2 Oct 2025)	LLM-based MAS topology adaptation	Dynamic graph designer that selects task-specific optimal graph configurations
MASFly (Liu et al., 14 Feb 2026)	Dynamic adaptation at test time	Retrieval-augmented SOP instantiation, Watcher agent, and personalized experience pool
Unified-MAS (Lin et al., 23 Mar 2026)	Knowledge-intensive automatic MAS	Offline node synthesis with Search-Based Node Generation and Reward-Based Node Optimization
MAS$O$3 (Wang et al., 29 Sep 2025)	Self-generative, self-configuring, self-rectifying MAS	Generator-implementer-rectifier tri-agent team with Collaborative Tree Optimization
GenoMAS (Liu et al., 28 Jul 2025)	Scientific discovery via gene-expression analysis	Guided-planning with typed message passing across six specialized LLM agents

Two further systems sharpen the broader pattern. “From Intent to Execution” describes an automated framework that replaces manual planning, manual agent selection, and manual execution-graph creation with an LLM-derived planner, dynamic call graph, orchestrator, agent recommender, and critique agent; the critique agent holistically reevaluates both agent and tool recommendations against the overall plan (Athrey et al., 5 May 2026). AutoMaAS, by contrast, frames the problem as self-evolving multi-agent architecture search with automatic operator generation, fusion, and elimination, dynamic cost-aware optimization, online feedback integration, and decision tracing mechanisms (Ma et al., 3 Oct 2025).

This literature does not define AutoMAS identically. Some systems emphasize orchestration hierarchy, some emphasize graph search, some emphasize node synthesis, and some emphasize benchmarking or scientific rigor. What unifies them is the relocation of design effort from manually authored pipelines to automated planning, retrieval, selection, supervision, or rectification.

5. Recurrent architectural motifs

A first recurring motif is hierarchical decomposition. BMW Agents structures the lifecycle of task automation through planning, execution, and verification, with a Planner Agent generating a DAG of subtasks and a dedicated Verifier Agent capable of triggering replanning (Crawford et al., 2024). Autonoma operationalizes a closely related separation of concerns through a Coordinator that validates intent, a Planner that generates structured workflows, and a Supervisor that dynamically manages specialized agents, reaching a reported 97% task completion rate and 98% successful agent handoff rate (Reda et al., 27 Feb 2026).

A second motif is adaptive structure selection. AMAS argues that conventional MAS architectures are restricted by inflexible, hand-crafted graph topologies, and answers this with a dynamic graph designer that ranks candidate collaboration graphs using a lightweight LoRA-adapted LLM (Leong et al., 2 Oct 2025). MASFly pushes the adaptation point to deployment time through retrieval-augmented SOP instantiation and a Watcher agent that monitors behaviors with reference to a personalized experience pool, achieving a 61.7% success rate on TravelPlanner (Liu et al., 14 Feb 2026). MAS$O$4 criticizes “generate-once-and-deploy” pipelines and instead lets a generator-implementer-rectifier meta-team dynamically compose and adapt a target MAS during execution (Wang et al., 29 Sep 2025).

A third motif is the decoupling of orchestration from domain expertise. Unified-MAS argues that knowledge-intensive domains are bottlenecked either by static libraries of general nodes or by on-the-fly node generation that couples domain logic to topology optimization. Its answer is offline node synthesis, in which Search-Based Node Generation retrieves external open-world knowledge and Reward-Based Node Optimization refines bottleneck nodes using a perplexity-guided reward, yielding up to a 14.2% gain while significantly reducing costs (Lin et al., 23 Mar 2026). MedMASLab makes an analogous intervention at the evaluation layer by decoupling data ingestion and modality-specific extraction from agent logic, standardizing 11 heterogeneous MAS architectures across 24 medical modalities, 11 organ systems, and 473 diseases (Qian et al., 10 Mar 2026).

A fourth motif is explicit supervision and self-correction. The critique agent in the intent-to-execution framework reevaluates plan-wide coherence and task-local fit (Athrey et al., 5 May 2026). GenoMAS routes code generation through review and domain-advice loops using typed message protocols, and its guided-planning framework lets programming agents advance, revise, bypass, or backtrack while operating over Action Units (Liu et al., 28 Jul 2025). AutoMaAS generalizes self-correction to the operator population itself through health-based operator generation, fusion, and elimination, alongside online feedback integration and counterfactual decision tracing (Ma et al., 3 Oct 2025).

Taken together, these mechanisms suggest that the contemporary meaning of AutoMAS is less “many agents working together” than “systems that can decide how many agents, which roles, which topology, which tools, and when to revise those choices.”

6. Limitations, evaluation tensions, and future directions

The wireless-network AutoMAS has three explicit limitations. First, current LLMs struggle with deep physical-layer reasoning and complex derivations, which requires careful curation of agent memory and external knowledge or tools. Second, decision time and computation cost for multi-agent orchestration and code synthesis are not fully optimized. Third, the evaluation focuses on single-user and RRH settings rather than multi-user, interference-limited scenarios (Yuan et al., 23 Nov 2025). The paper correspondingly identifies future work on optimizing decision latency and computational overhead, extending to multi-user interference-limited environments, supporting joint accuracy-latency tradeoffs, and exploring new agent roles such as semantic communication and network slicing (Yuan et al., 23 Nov 2025).

These limitations resonate with broader AutoMAS research. MedMASLab identifies fragmented architectures, non-uniform data ingestion pipelines, inconsistent visual-reasoning evaluation, and a lack of cross-specialty benchmarking in medical MAS; its results further indicate that current architectures exhibit significant fragility when transitioning between specialized medical sub-domains (Qian et al., 10 Mar 2026). AMAS diagnoses the rigidity of fixed collaboration graphs, while MASFly argues that many systems lack dynamic adaptability after deployment (Leong et al., 2 Oct 2025, Liu et al., 14 Feb 2026). Unified-MAS describes a severe architectural coupling between domain-specific node generation and high-level topology optimization in knowledge-intensive domains (Lin et al., 23 Mar 2026). MAS$O$5 frames the dominant failure mode as the brittleness of generate-once-and-deploy pipelines under real-world dynamism and uncertainty (Wang et al., 29 Sep 2025).

Evaluation itself has become a central research problem. MedMASLab’s automated clinical reasoning evaluator rejects exact-match style assessment in favor of zero-shot semantic judging with a large vision-LLM (Qian et al., 10 Mar 2026). GenoMAS evaluates preprocessing with Composite Similarity Correlation and downstream discovery with $O$6, reporting 89.13% for data preprocessing and 60.48% for gene identification on GenoTEX (Liu et al., 28 Jul 2025). The intent-to-execution framework measures planning, agent selection, and task completion jointly, emphasizing recall, robustness, and structured workflow correctness rather than tool retrieval alone (Athrey et al., 5 May 2026). A plausible implication is that AutoMAS research is moving from anecdotal demonstrations toward multi-axis evaluation of correctness, cost, latency, robustness, and semantic fidelity.

More generally, the field appears to be converging on three unresolved questions. The first is how to combine dynamic adaptability with domain guarantees, which the wireless AutoMAS addresses by selecting among theoretically grounded algorithms rather than replacing them. The second is how to scale automatic generation into knowledge-intensive domains without hallucinated node logic, motivating retrieval-grounded or offline-synthesized expertise. The third is how to verify or assure these systems, a question that reconnects the modern automatic-MAS literature to the earlier AutoMAS workshop emphasis on formal specification, modeling, and verification of autonomous systems (Luckcuck et al., 2020).

In that sense, AutoMAS now denotes a research trajectory rather than a single architecture. Its concrete wireless realization demonstrates environment-driven algorithm self-adaptation; its broader literature seeks automatic generation, supervision, and rectification of multi-agent workflows; and its open problems increasingly span algorithmic soundness, domain grounding, cost-performance optimization, and assurance.