WirelessAgent++: Automated Agentic Workflow Design and Benchmarking for Wireless Networks

Jingwen Tong; Zijian Li; Fang Liu; Wei Guo; Jun Zhang

cs.NIFeb 28, 2026

WirelessAgent++: Automated Agentic Workflow Design and Benchmarking for Wireless Networks

Authors: Jingwen Tong, Zijian Li, Fang Liu, Wei Guo, Jun Zhang

Organizations: College of Electronics and Information Engineering, Shenzhen University, Shenzhen, China · Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR

Abstract

The integration of large language models (LLMs) into wireless networks has sparked growing interest in building autonomous AI agents for wireless tasks. However, existing approaches rely heavily on manually crafted prompts and static agentic workflows, a process that is labor-intensive, unscalable, and often suboptimal. In this paper, we propose WirelessAgent++, a framework that automates the design of agentic workflows for various wireless tasks. By treating each workflow as an executable code composed of modular operators, WirelessAgent++ casts agent design as a program search problem and solves it with a domain-adapted Monte Carlo Tree Search (MCTS) algorithm. Moreover, we establish WirelessBench, a standardized multi-dimensional benchmark suite comprising Wireless Communication Homework (WCHW), Network Slicing (WCNS), and Mobile Service Assurance (WCMSA), covering knowledge reasoning, code-augmented tool use, and multi-step decision-making. Experiments demonstrate that \wap{} autonomously discovers superior workflows, achieving test scores of $78.37\%$ (WCHW), $90.95\%$ (WCNS), and $97.07\%$ (WCMSA), with a total search cost below $\$ 5 $per task. Notably, our approach outperforms state-of-the-art prompting baselines by up to$ 31% $and general-purpose workflow optimizers by$ 11.1%$, validating its effectiveness in generating robust, self-evolving wireless agents. The code is available at https://github.com/jwentong/WirelessAgent-R2.

Figures & tables

Fig. 1: System overview: from WirelessAgent (manual design) to WirelessAgent ++ (automated design). Left: in WirelessAgent, a human expert iteratively designs a fixed agentic workflow through multi-round dialogue with an LLM. Right: in WirelessAgent ++ , an Optimizer LLM jointly searches over the workflow structure and tool-calling strategy; the resulting workflow is executed by an Executor LLM on WirelessBench , automatically producing distinct, task-adaptive workflows (e.g., Q&A, network slicing (NS), and mobile service assurance (MSA)) without manual engineering.

Fig. 2: The overall architecture of WirelessAgent ++ , comprising three components. Search Space: each workflow node contains fixed parameters and optimizable prompts, composed from a library of operators and domain-specific tools. Search Algorithm: an MCTS-based optimizer iterates through four phases: penalized Boltzmann probability selection, LLM-based expansion, executing evaluation, and experience backpropagation, using an advanced Optimizer LLM for expansion and a cost-efficient Executor LLM for evaluation. WirelessBench : example task-adaptive workflows automatically discovered for WCHW, WCNS, and WCMSA, each exhibiting distinct operator compositions and tool-calling patterns.

Parameter	Value
TX power	30 dBm
Carrier frequency	2.4 GHz
Bandwidth	20 MHz
Noise figure	8 dB
Thermal noise	$-174+10\log_{10}(\text{BW})$ dBm
User height	1.5 m

TABLE I: Ray-tracing channel predictor parameters.

Level	Condition	Allowed Changes
Ultra-conservative	Saturated	Single word / punctuation only
Conservative	$s\geq 0.65$	Minor prompt wording; no structural changes
Moderate	$0.50\leq s<0.65$	Single structural change; no new branches
Aggressive	$s<0.50$	Structural changes, new operators, ToolAgent additions

TABLE II: Heuristic critic aggressiveness levels and allowed mutation types. A node is saturated if it has

\geq 2

failed mutations and zero successes.

Condition	Score
Relative error $<1\%$	1.0
Relative error $<5\%$	0.9
Relative error $<10\%$	0.7
Off by $\times 10^{3}$ or $\times 10^{6}$ (unit error)	0.5
Off by $\times 2$ (common miscalculation)	0.3
Otherwise	0.0

TABLE III: WCHW numeric scoring tiers. The system also detects common errors such as factor-of-

10^{3}

and factor-of-

2

mistakes.

Sub-Metric	Weight	Scoring Rule
Slice type	25%	Binary: exact match $\to$ 1.0, else 0.0
CQI	15%	Exact $\to$ 1.0; $\pm 1\to$ 0.8; $\ldots$ ; $\pm 3{\text{--}}4\to$ 0.2
Bandwidth	35%	Rel. err.: ${<}2\%{\to}1.0$ ; $\ldots$ ; ${<}20\%{\to}.70$
Throughput	25%	Same tiered scheme as bandwidth

TABLE IV: WCNS composite scoring: 4-metric weighted scheme.

Sub-Metric	Weight	Scoring Rule
Position prediction	15%	Distance decay: $\max(0,1-(d/20)^{1.2})$
CQI prediction	15%	Exact $\to$ 1.0; $\ldots$ ; $\pm 3\to$ 0.50
Slice type	20%	Binary: exact $\to$ 1.0, else 0.0
Bandwidth allocation	25%	Same tiered scheme as WCNS
Throughput	20%	Same tiered scheme as WCNS
QoS verification	5%	Binary: correct Yes/No $\to$ 1.0, else 0.0

TABLE V: WCMSA composite scoring: 6-metric weighted scheme.

Benchmark	Val.	Test	Metric	Task Type
WCHW	348	1,044	Accuracy	Knowledge reasoning
WCNS	250	750	Composite	Code + tool use
WCMSA	250	750	Composite	Multi-step diagnosis

TABLE VI: Statistics of the WirelessBench benchmark suite.

				WirelessBench (Solve Rate)
Method	HotpotQA (F1)	DROP (F1)	MATH (Solve Rate)	Before	After
Qwen-turbo-latest [ 31 ]	0.1423	0.5641	0.6412	0.5013	0.5834
CoT [ 32 ]	0.5379	0.7607	0.6828	0.5032	0.6032
MedPrompt [ 33 ]	0.5411	0.7559	0.6996	0.5422	0.6122
ADAS [ 13 ]	0.5110	0.7423	0.4953	0.4813	0.5313
AFlow [ 20 ]	0.5823	0.7811	0.7864	0.5152	0.6992

TABLE VII: Performance comparison on general NLP benchmarks and the wireless benchmark. WirelessBench “Before” uses the original dataset; “After” applies the psychometric data-cleaning pipeline (Section V-B ) that detects and corrects low-quality samples. Best results in bold .

Fig. 3: Method comparison on the test dataset of the WCHW benchmark. WirelessAgent ++ outperforms the AFlow and all prompting baselines.

Fig. 4: Workflow evolution of WirelessAgent ++ on the validation dataset of the WCHW benchmark.

Panel of Figure 12 — Fig. 5: WCNS case study. (a) Overall composite score comparison across methods. (b) Per-metric breakdown: WirelessAgent ++ achieves $\geq$ 89% on all four sub-metrics. (c) Workflow evolution on the validation set showing the three-phase optimization trajectory.

Panel of Figure 13 — Fig. 6: Total throughput and bandwidth idle rate versus the number of users under three HKUST campus scenarios [ 12 ] . WirelessAgent ++ consistently achieves the highest throughput among all LLM-based methods.

Panel of Figure 14 — Fig. 7: WCMSA case study. (a) Overall composite score comparison across methods. (b) Per-metric breakdown: WirelessAgent ++ achieves $\geq$ 96% on all six sub-metrics. (c) Workflow evolution on the validation set showing the three-phase optimization trajectory.

Configuration	Accuracy (%)	$\boldsymbol{\Delta}$
WirelessAgent ++ (Full)	81.78	–
w/o Penalized Selection	79.11	$-2.67$
w/o Heuristic Critic	80.34	$-1.44$
w/o 3-Class Experience	78.59	$-3.19$
w/o Domain Tools	62.44	$-19.34$
AFlow (No enhancements) †	69.92	$-11.86$

TABLE VIII: Ablation study on WCHW validation set.

\Delta

denotes the change from the full WirelessAgent ++ configuration.

	WCHW	WCNS	WCMSA
Search Rounds	19	11	11
Wall-Clock Time (min)	63	13	14
Total Search Cost (USD)	4.95	0.99	1.05
Per-Problem Cost (USD)	0.00083	0.00056	0.00068

TABLE IX: MCTS search cost and inference cost across benchmarks.

CQI	Modulation	Code Rate $\times 1024$	$\boldsymbol{\eta}$ (bps/Hz)
1	QPSK	78	0.15
2	QPSK	120	0.23
3	QPSK	193	0.38
4	QPSK	308	0.60
5	QPSK	449	0.88
6	QPSK	602	1.18

TABLE X: Complete CQI-to-spectral-efficiency mapping (3GPP TS 38.214, Table 5.2.2.1-2).

Operator	Description
Custom	Invokes an LLM with input and instruction prompt; the most flexible operator
ToolAgent	ReAct-based agent (Algorithm 1 ) with tool calling; max $I$ iterations with early stopping
CodeLevel	Deterministic, LLM-free tool execution via compiled code (e.g., CodeLevelRayTracing , CodeLevelKalmanPredictor ); zero variance, near-zero cost
ScEnsemble	Self-consistency ensemble [ 34 ] : maps $N$ candidate solutions to letters and selects the most consistent answer
MdEnsemble	Medicine-inspired ensemble [ 33 ] : shuffles $N$ answers, votes 5 times, returns majority answer
Programmer	Generates and executes Python code (3 retries, 30s timeout per attempt)

TABLE XI: Complete operator repertoire available to WirelessAgent ++ .

Service	Slice	Min Rate (Mbps)
4K video streaming	eMBB	25
HD movie download	eMBB	50
Cloud gaming (high graphics)	eMBB	35
AR navigation (3D maps)	eMBB	20
VR social experience	eMBB	40
8K live sports broadcast	eMBB	80

TABLE XII: WCMSA service types and minimum QoS requirements.

Region	OSM File	Typical Bldg.	Propagation
North	HKUST_North.osm	Low-rise academic	LOS-dominant
Center	HKUST_Center.osm	Mixed-height	Moderate NLOS
South	HKUST_South.osm	High-rise residential	Deep NLOS

TABLE XIII: HKUST campus ray-tracing regions used in WirelessBench .

Parameter	Default	Description
$K$ (top- $K$ )	5	Parent candidates for Boltzmann selection
$\lambda$	0.3	Uniform exploration weight in ( 11 )
$\alpha$	0.2	Boltzmann temperature in ( 11 )
$\epsilon$	0.02	3-class significance threshold in ( 13 )
$V$	5	Validation runs per evaluation
$T$	20	Maximum optimization rounds

TABLE XIV: Complete hyperparameter summary for WirelessAgent ++ .

Wireless Local Area Networks

See all 18 papers

Apr 11, 2026

The AI Telco Engineer: Toward Autonomous Discovery of Wireless Communications Algorithms

Agentic AI is rapidly transforming the way research is conducted, from prototyping ideas to reproducing results found in the literature. In this paper, we explore the ability of agentic AI to autonomously design wireless communication algorithms. To that end, we implement a dedicated framework that leverages large language models (LLMs) to iteratively generate, evaluate, and refine candidate algorithms. We evaluate the framework on three tasks spanning the physical (PHY) and medium access control (MAC) layers: statistics-agnostic channel estimation, channel estimation with known covariance, and link adaptation. Our results show that, in a matter of hours, the framework produces algorithms that are competitive with and, in some cases, outperforming conventional baselines. Moreover, unlike neural network-based approaches, the generated algorithms are fully explainable and extensible. This work represents a first step toward the autonomous discovery of novel wireless communication algorithms, and we look forward to the progress our community makes in this direction.

Fayçal Aït Aoudia, Jakob Hoydis, Sebastian Cammerer +3

All authors are with NVIDIA.

Automated Algorithm Discovery Wireless Communications Large Language Model-Guided Scientific Discovery Channel Estimation

Ask PDF

Mar 2, 2026

Federated Agentic AI for Wireless Networks: Fundamentals, Approaches, and Applications

Agentic artificial intelligence (AI) presents a promising pathway toward realizing autonomous and self-improving wireless network services. However, resource-constrained, widely distributed, and data-heterogeneous nature of wireless networks poses significant challenges to existing agentic AI that relies on centralized architectures, leading to high communication overhead, privacy risks, and non-independent and identically distributed (non-IID) data. Federated learning (FL) has the potential to improve the overall loop of agentic AI through collaborative local learning and parameter sharing without exchanging raw data. This paper proposes new federated agentic AI approaches for wireless networks. We first summarize fundamentals of agentic AI and mainstream FL types. Then, we illustrate how each FL type can strengthen a specific component of agentic AI's loop. Moreover, we conduct a case study on using FRL to improve the performance of agentic AI's action decision in low-altitude wireless networks (LAWNs). Finally, we provide a conclusion and discuss future research directions.

Lingyi Cai, Yu Zhang, Ruichen Zhang +5

Research Center of 6G Mobile Communications, School of Cyber Science and Engineering, Huazhong University of Science and Technology, Wuhan, 430074, China · College of Computing and Data Science, Nanyang Technological University, Singapore · School of Engineering, Edith Cowan University, Perth, WA 6027, and the School of Computer Science and Engineering, University of New South Wales (UNSW), Sydney, NSW 2033, Australia +2

Wireless Communications Agentic RL Federated RL

Ask PDF

Oct 16, 2026

LoopHIL: Looping Hardware into LLM-assisted Agents for Wireless Embedded Systems

LLM agents are widely used for software development and debugging, but their adoption in embedded systems, especially for wireless tasks, remains limited. This is because embedded development requires hardware-in-the-loop testing. Agents cannot directly observe the spectrum, probe pins, or measure power, so they cannot verify wireless link functionality; currently, human engineers perform this verification. We systematically evaluate whether LLM agents can establish wireless links when provided with hardware access. We introduce LoopHIL, a framework that lets agents control a multi-device RF bench, monitor it with a spectrum analyzer, logic analyzer, and energy meter, and validate links against a verification ladder, all within code-enforced safety limits. Across four increasingly complex tasks, from generating a carrier to backscatter decoding by a commercial receiver, and using six models from two vendors, hardware-in-the-loop increases task completion rates from 54% to 79% and from 17% to 83% on the most challenging task. It also halves task time and output tokens. With debugging tools, agents decrease tag power consumption by up to 96%. However, small models often report unverified claims, so independent hardware checks remain necessary.

Dhairya Shah, Spanddhana Sara, Savitha Viswanadh Kandala +1

National University of Singapore, Singapore

LLM Agent Evaluation LLM Agent Harnesses LLM Agent Reliability Internet of Things

Ask PDF

Abstract

Wireless Local Area Networks

Explore similar work

The AI Telco Engineer: Toward Autonomous Discovery of Wireless Communications Algorithms

Federated Agentic AI for Wireless Networks: Fundamentals, Approaches, and Applications

LoopHIL: Looping Hardware into LLM-assisted Agents for Wireless Embedded Systems