WirelessAgent++: Automated Wireless Workflows

• WirelessAgent++ is an agent-design framework for wireless networks that automates the synthesis of executable, modular workflows for diverse tasks.
• It employs a domain-adapted Monte Carlo Tree Search and a constrained operator library to optimize control-flow, tool integration, and workflow structure.
• Empirical evaluations on WirelessBench benchmarks demonstrate high performance and cost efficiency, outperforming traditional prompt-based methods.

WirelessAgent++ is an agent-design framework for wireless networks that automates the synthesis of agentic workflows for heterogeneous wireless tasks by treating each workflow as executable code composed of modular operators and solving the resulting program search problem with a domain-adapted Monte Carlo Tree Search (MCTS) algorithm. It is coupled with WirelessBench, a standardized multi-dimensional benchmark suite spanning Wireless Communication Homework (WCHW), Network Slicing (WCNS), and Mobile Service Assurance (WCMSA). In the reported evaluation, automatically discovered workflows achieve test scores of $78.37\%$ on WCHW, $90.95\%$ on WCNS, and $97.07\%$ on WCMSA, while keeping total search cost below $\$5$per task and per-problem inference cost below$\$0.001$$(<a href="/papers/2603.00501" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Tong et al., 28 Feb 2026</a>).</p> <h2 class='paper-heading' id='conceptual-scope-and-motivation'>1. Conceptual scope and motivation</h2> <p>WirelessAgent++ addresses a practical bottleneck in applying LLMs to wireless tasks: manually crafted prompts and static agentic workflows do not scale across heterogeneous task families and are often suboptimal. In this formulation, manual agent engineering—deciding tool sequences, control flow, error handling, and formatting—becomes a new feature-engineering and pipeline-engineering burden as tasks diversify and models evolve (<a href="/papers/2603.00501" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Tong et al., 28 Feb 2026</a>).</p> <p>The framework therefore shifts the optimization target from model parameters to workflow structure. Rather than assuming a fixed ReAct-style prompt or a fixed sequence of tools, WirelessAgent++ jointly optimizes control-flow topology, tool-use strategy, and prompts. A workflow is represented as an executable program$$W$ with a typed <code>__call__</code> interface returning $(\hat{y}, c)$, where$\hat{y}$is the answer and$c$$is the accumulated <a href="https://www.emergentmind.com/topics/adversarial-prompt-injection-api" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">API</a> or tool cost. Search is conducted over short, structured programs of roughly 10–50 lines with clearly defined inputs and outputs (<a href="/papers/2603.00501" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Tong et al., 28 Feb 2026</a>).</p> <p>This design places WirelessAgent++ in the category of automated agent designers rather than individual hand-built agents. A plausible implication is that the system is intended to remain useful as models, tools, and task distributions change, because the optimized object is the workflow graph rather than a single handcrafted prompt.</p> <h2 class='paper-heading' id='workflow-representation-and-operator-system'>2. Workflow representation and operator system</h2> <p>The search space of WirelessAgent++ is built from a constrained operator library. Core operator types include <code>Custom(x, p)</code> for LLM invocation with instruction prompt$$90.95\%$$0, <code>ToolAgent(x, s)</code> for ReAct-based closed-loop tool use, <code>CodeLevel(x, f)</code> for deterministic tool execution, <code>ScEnsemble({y_i}, x)</code> for self-consistency voting, and modules such as <code>Review</code>, <code>Revise</code>, <code>Programmer</code>, <code>Test</code>, <code>AnswerGenerate</code>, <code>Format</code>, and <code>AnswerValidator</code> for robust solution construction, code generation and evaluation, and format validation (<a href="/papers/2603.00501" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Tong et al., 28 Feb 2026</a>).</p> <p>Two operator modes are especially important. <code>ToolAgent</code> is an LLM-driven orchestration loop with up to$$90.95\%$1 Think–Act–Observe iterations, early stopping after two consecutive tool errors, strict XML parsing, and auto-correction for bracket, LaTeX, and JSON anomalies. CodeLevel instead compiles a tool call into deterministic execution with zero variance and near-zero cost. The search often discovers a useful ToolAgent pattern first and later replaces it with a stable CodeLevel operator to reduce variance and inference cost (Tong et al., 28 Feb 2026).

WirelessAgent++ also includes a wireless-specific tool library. The telecom formula retriever is a retrieval-augmented component over a curated set of 31 formulas with relevance score

$90.95\%$2

The precision calculator is SciPy-backed, supports special functions and BER libraries, implements Marcum Q with integration or series fallback, and provides 10-digit precision. The ray-tracing channel predictor uses OpenStreetMap geometry and LOS/NLOS detection, with

$90.95\%$3

$90.95\%$4

and CQI mapping

$90.95\%$5

The Kalman filter predictor uses a constant-velocity model

$90.95\%$6

with $90.95\%$7, $90.95\%$8, and $90.95\%$9 (Tong et al., 28 Feb 2026).

A common misconception is that WirelessAgent++ is simply a prompting scheme. The operator library and executable-program abstraction show that it is instead a structured workflow synthesis system with typed operators, explicit tool modes, validators, and deterministic compilation paths.

3. Program search and domain-adapted MCTS

WirelessAgent++ formulates workflow optimization for task $97.07\%$0 as

$97.07\%$1

subject to a practical search-cost budget $97.07\%$2. Budget control is enforced through a low-cost Executor LLM, median-of-$97.07\%$3 evaluation runs, and convergence detection for early stopping (Tong et al., 28 Feb 2026).

The MCTS loop contains three tailored components. First, selection uses penalized Boltzmann sampling rather than standard UCT/UCB. Given top-$97.07\%$4 parents $97.07\%$5, each node receives a history-aware exploration penalty

$97.07\%$6

with $97.07\%$7 if $97.07\%$8 and $97.07\%$9 otherwise, and the parent probability is

$\$0

where $\$1. This implements a soft-pruning strategy that downweights saturated or failure-heavy branches while preserving recovery potential (Tong et al., 28 Feb 2026).

Second, expansion is carried out by an Optimizer LLM that proposes a single focused code or prompt change, typically no more than five lines, guided by a critic report and formatted experience. Duplicated or previously harmful modifications are rejected. Third, backpropagation uses 3-class experience replay with significance threshold $\$2:

$\$3

Successes become positive exemplars, failures are blacklisted, and neutral outcomes are retained without penalty in order to avoid chasing noise (Tong et al., 28 Feb 2026).

Stopping relies on convergence detection over the running top-$\$4 scores. If $\$5 is the mean of the running top-$\$6 scores and $\$7, search halts when $\$8 for $\$9 consecutive rounds, with default patience $per task and per-problem inference cost below$0 and $per task and per-problem inference cost below$1 (Tong et al., 28 Feb 2026).

The implementation uses a two-tier LLM design: an Optimizer LLM such as Claude-Opus-4.5 or GPT-4o for mutation, and an Executor LLM such as Qwen-turbo-latest or DeepSeek-V3 for evaluation and runtime execution, all with temperature set to zero (Tong et al., 28 Feb 2026).

4. WirelessBench and canonical discovered workflows

WirelessBench is a standardized benchmark suite with deterministic ground truths covering three task families: knowledge reasoning, code-augmented tool use, and multi-step decision-making with mobility. The data pipeline combines authoritative sources, funnel-style psychometric cleaning, LLM-based augmentation followed by deterministic re-computation of ground truths, and human validation (Tong et al., 28 Feb 2026).

Benchmark	Task focus	Size
WCHW	Knowledge reasoning and numerics	348 val / 1,044 test
WCNS	Code-augmented tool use	250 val / 750 test
WCMSA	Multi-step decision-making with mobility	250 val / 750 test

WCHW contains 1,392 problems from textbooks and evaluates formulas, unit conversion, and special functions. Inputs are natural-language questions; outputs are structured numeric, text, or formula answers; scoring is multi-strategy and format-aware, including numeric relative error tiers, LaTeX formula similarity, and keyword-based text scoring. WCNS models a network-slicing scenario with eMBB and URLLC slices, user position, and service intent; it requires CQI prediction via ray-tracing and proportional-fairness bandwidth allocation, and outputs slice type, CQI, bandwidth, and throughput. WCMSA extends WCNS with mobility, requiring trajectory prediction by Kalman filtering, CQI estimation at the predicted position, and QoS verification; outputs include predicted position, CQI, slice, bandwidth, throughput, and QoS yes/no (Tong et al., 28 Feb 2026).

The benchmark is paired with task-specific discovered workflows. On WCHW, the canonical pipeline is Custom → ToolAgent, described as a Reason-then-Verify pattern: the first stage solves with domain formulas, and the second generates Python for verification and unit normalization before structured extraction. On WCNS, the discovered pattern is CodeLevelRayTracing → Custom, a Tool-then-Reason pipeline in which deterministic CQI is injected before intent classification and resource computation with CQI-to-$per task and per-problem inference cost below$2 lookup. On WCMSA, the discovered pattern is CodeLevelKalmanPredictor → CodeLevelRayTracing → Custom, described as Predict–Estimate–Reason: future position is predicted, future CQI is estimated deterministically, and the final step performs slice decision, bandwidth allocation, throughput calculation, and QoS verification (Tong et al., 28 Feb 2026).

These three workflows illustrate a central design claim of WirelessAgent++: task families with different structure call for different operator graphs rather than a single universal prompt.

5. Empirical performance, efficiency, and ablations

On test sets, WirelessAgent++ reports $per task and per-problem inference cost below$3 on WCHW, $per task and per-problem inference cost below$4 on WCNS, and $per task and per-problem inference cost below$5 on WCMSA; on validation trajectories it reaches $per task and per-problem inference cost below$6, $per task and per-problem inference cost below$7, and $per task and per-problem inference cost below$8, respectively. Search cost remains below %%%%49$0.001$50%%%%\$0.001$1 for WCHW over 19 rounds in roughly 63 minutes, $0.001$20.99$0.001$3\$0.001$4 for WCMSA over 11 rounds in roughly 14 minutes. Per-problem inference cost remains below $0.001$50.001$0.001$6\$0.001$7 reported for WCHW (<a href="/papers/2603.00501" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Tong et al., 28 Feb 2026</a>).</p> <div class='overflow-x-auto max-w-full my-4'><table class='table border-collapse w-full' style='table-layout: fixed'><thead><tr> <th>Benchmark</th> <th>Test score</th> <th>Representative search cost</th> </tr> </thead><tbody><tr> <td>WCHW</td> <td>78.37%</td> <td>\$4.95 WCNS 90.95% \$0.99 WCMSA 97.07% \$1.05

Relative to baselines, the framework outperforms prompting baselines such as MedPrompt or CoT by up to 31 percentage points and exceeds general-purpose workflow optimizers such as AFlow by 11.1 percentage points. In WCNS, tool integration yields a dramatic CQI prediction jump, with CQI accuracy around $0.001$8 versus around $0.001$9 for prompt-only configurations (Tong et al., 28 Feb 2026).

Ablation results attribute the largest effect to domain tools: removing them causes a $W$0 percentage-point drop on WCHW validation. Disabling 3-class replay reduces performance by $W$1 points, removing penalized selection reduces it by $W$2 points, and removing the heuristic critic reduces it by $W$3 points. Sensitivity analysis on $W$4 shows that overly small values induce noise chasing, while $W$5 balances stability and improvement detection. Default hyperparameters reported to generalize across benchmarks are top-$W$6, $W$7, $W$8, $W$9, and $(\hat{y}, c)$0 (Tong et al., 28 Feb 2026).

The empirical explanation offered by the framework is architectural rather than purely model-centric. ToolAgent supports closed-loop discovery of useful tool chains, later compiled into CodeLevel operators; penalized Boltzmann selection preserves exploration while softly pruning failure-saturated branches; median-of-$(\hat{y}, c)$1 scoring and 3-class replay protect against stochasticity and noisy metrics; and the two-tier Optimizer/Executor split controls search cost without sacrificing mutation diversity (Tong et al., 28 Feb 2026).

6. Relation to adjacent agentic wireless research and stated limitations

WirelessAgent++ emerges within a broader body of work that applies agentic AI to wireless systems, but its distinctive contribution is automated workflow search over executable operator graphs. Earlier WirelessAgent work framed wireless agents around perception, memory, planning, and action, implemented a LangGraph-based network-slicing pipeline, and reported $(\hat{y}, c)$2 higher bandwidth utilization than a Prompt-based method while remaining only $(\hat{y}, c)$3 below Rule-based optimality (Tong et al., 2 May 2025). ComAgent, by contrast, organized a closed-loop Perception–Planning–Action–Reflection cycle with specialized Literature, Planning, Coding, and Scoring agents, achieving a 100.00% Problem Formulation Rate and 100.00% Code Execution Rate across 25 wireless tasks, with a 72.00% Solution Solved Rate (Li et al., 27 Jan 2026). AutoMAS emphasized environment-aware algorithm selection among theoretically grounded wireless solvers (Yuan et al., 23 Nov 2025), while RadioMaster targeted autonomous radio signal generation through domain retrieval, multi-agent role specialization, and emulator-gated verification (Lei et al., 1 Jun 2026). Other related directions include secure and energy-efficient supervisor–agent networks with friendly jamming (Song et al., 16 Feb 2026), MCP-based Internet-of-Experts environment-aware LLM agents (Liu et al., 3 May 2025), terminal-side personal agents with offline reflection and deterministic online execution (Tan et al., 22 Jun 2026), intent-aware TinyML handover systems (Saleh et al., 2 Aug 2025), semantic-aware wireless agent networks under ILAC (Zhao et al., 1 Apr 2026), and agent-native wireless architectures built around O-RAN programmability and loop-level interface semantics (Liu et al., 15 May 2026).

Within this landscape, WirelessAgent++ is explicitly limited to single-agent optimization; multi-agent coordination across cells and RAN tiers remains open. The tool library is fixed during search, so dynamic tool discovery and composition are not yet supported. Structural workflows transfer across models, but prompt tuning may be executor-dependent, making model-agnostic prompt strategies an open direction. The wireless models in WirelessBench are simplified: single-cell ray tracing is used, with no inter-cell interference, no MIMO, and no RIS. Extending the benchmark to richer PHY/MAC settings such as beamforming and interference coordination is identified as future work. Hyperparameter calibration, especially for $(\hat{y}, c)$4 and critic thresholds, may require retuning under different score distributions. Finally, although deterministic tools mitigate hallucinations, deployment is stated to require safeguards for incorrect tool inputs, out-of-distribution environments, and human-in-the-loop review for critical operations such as URLLC healthcare (Tong et al., 28 Feb 2026).

A recurring misconception is that the framework solves wireless control by replacing domain methods with unconstrained language-model reasoning. The reported design suggests the opposite: WirelessAgent++ depends on deterministic tools, typed workflow operators, validation logic, robust scoring, and a benchmark with deterministic ground truths. In that sense, it represents an overview of LLM-based agent planning with domain-specific execution, not a displacement of wireless-domain structure.