# WirelessAgent-R2: MCTS-Driven Agent Workflow Search for Wireless Tasks

> WirelessAgent-R2 (branded WirelessAgent++) automates the design of LLM agent workflows for wireless communication problems using Monte Carlo Tree Search, reporting up to 31% accuracy gains over hand-crafted baselines on its WirelessBench suite.

**jwentong/WirelessAgent-R2** — WirelessAgent-R2 is an automated workflow optimization system that uses Monte Carlo Tree Search (MCTS) to iteratively discover and refine LLM-based problem-solving strategies for wireless communication tasks, achieving a 20% performance improvement from 0.583 to 0.78 accuracy through intelligent exploration of prompt designs and operator

- Repository: https://github.com/jwentong/WirelessAgent-R2
- Stars: 23 · Forks: 2
- Language: Python
- License: MIT
- Published: 2026-08-24 · Updated: 2026-08-24 · Language: en
- Canonical page: https://hysenlabs.com/projects/jwentong-wirelessagent-r2

## What WirelessAgent-R2 Solves and Who It Is For

WirelessAgent-R2, presented in the README as WirelessAgent++, targets a specific pain point: designing LLM-based agent workflows for wireless network problems is manual, iterative, and slow. A human expert has to craft prompts, decide which tools to call, and chain operations by hand. The project automates that design process by treating workflow construction as a program search problem. It uses Monte Carlo Tree Search (MCTS) to explore combinations of operators, prompts, and tool-calling patterns, then evaluates them on a benchmark suite called WirelessBench. The intended audience is researchers or engineers working on wireless communication tasks such as knowledge reasoning, network slicing, and mobile service assurance, who want to replace manual prompt engineering with an automated search. The README claims the system outperforms hand-crafted baselines by up to 31 percentage points and beats a general-purpose workflow optimizer by 11.1 points, though those numbers come from the project's own reporting and should be treated as claims, not independent results.

## The MCTS Mechanism and Domain-Specific Enhancements

The core mechanism is an optimizer LLM that searches over workflow structures and tool-calling strategies via MCTS. The README names Claude-Opus-4.5 as the optimizer and Qwen-turbo-latest as the executor, but those are specific model choices, not hard requirements. The search space is composed of five operator types: Custom for direct LLM calls, ToolAgent for ReAct-style interleaved reasoning and tool calls, CodeLevel for deterministic, LLM-free tool execution, ScEnsemble for self-consistency voting, and AnswerGenerate for structured final answers. Three enhancements to standard MCTS are described. Penalized Boltzmann Selection applies a temperature-controlled penalty to visited nodes to avoid getting stuck in poor subtrees. A Maturity-Aware Heuristic Critic uses a lightweight LLM to pre-screen mutations before full evaluation, cutting down on expensive evaluations. Three-Class Experience Replay classifies mutation outcomes as Success, Neutral, or Failure instead of binary, which the authors argue prevents noise from corrupting the experience buffer. These are concrete algorithmic choices, though the README does not provide pseudocode or hyperparameter values, so reproducibility depends on the source code.

## Key Finding: Compiling ToolAgents into CodeLevel Pipelines

A significant design insight reported in the README is that the MCTS optimizer discovers that ToolAgent-based workflows can be compiled into CodeLevel pipelines. This means the LLM is removed from the tool-calling path, replaced by deterministic execution, while maintaining accuracy. The workflow evolution examples illustrate this. In WCNS, a seed workflow at 61.3% accuracy jumps to 90.5% after ToolAgent discovers a ray-tracing tool, then to 92.18% after compiling to CodeLevelRayTracing. In WCMSA, the seed at 65.76% rises to 93.59% after multi-tool discovery (Kalman filter and ray-tracing), then to 96.89% after compilation. The pattern is consistent: tool discovery gives the biggest jump, and compilation adds a smaller but non-trivial gain while reducing variance and cost. This is a practical advantage because CodeLevel operators are described as having zero variance and near-zero cost, which matters for production deployment where LLM calls add latency and expense.

## Getting It Running: Commands and Configuration from the README

The README includes a Quick Start section, but the cleaned version provided does not contain the actual commands or configuration keys. It only lists the table of contents and project structure references. Based on the repository layout, the project is Python-based with an MIT license. The README mentions a citation section and a license file, but no installation instructions are visible in the provided material. The absence of concrete commands is a limitation for this review. To get it running, you would need to clone the repository from GitHub and inspect the source for entry points, likely a main script or a configuration file that specifies the optimizer and executor model names. The README does mention that the full optimization search costs under $5 per task and per-problem inference costs under $0.001, but those are cost estimates, not setup steps. If you are evaluating this project, expect to read the code to figure out how to invoke the MCTS search on your own tasks.

## WirelessBench: The Benchmark Suite and Its Construction

WirelessBench is a standardized benchmark suite with 3,392 problems across three dimensions. WCHW (Wireless Communication Homework) has 1,392 problems focused on knowledge reasoning, including multi-step formula application and unit conversion. WCNS (Network Slicing) has 1,000 problems that require code and tool use, specifically ray-tracing CQI prediction followed by bandwidth allocation. WCMSA (Mobile Service Assurance) has 1,000 problems involving multi-step decisions, such as Kalman prediction, CQI estimation, and QoS assurance. The data construction pipeline is described in detail: seed problems from wireless textbooks (Goldsmith, Molisch) and 3GPP/IEEE standards, then a 10-LLM funnel pipeline for psychometric cleaning using item-total correlation, Mokken scale analysis, and inter-item consistency. LLM-based augmentation generates problem text, but all ground truths are computed by deterministic solvers, which is a strong design choice to avoid LLM hallucination in answers. Human validation by graduate students verifies every problem. This pipeline is a genuine contribution, but the README does not state whether the benchmark data is included in the repository or requires separate download.

## Reported Results and the Cost of Optimization

The README reports optimization results from 19 rounds on WCHW with a total search cost of $4.95 and about 63 minutes of wall-clock time. The validation score improved from 62.44% at Round 1 to 81.78% at Round 14, a gain of 19.34 percentage points. The largest single gain came at Round 2 with tool integration, adding 18.42 points. Held-out test scores are listed separately: 78.37% for WCHW, 90.95% for WCNS, and 97.07% for WCMSA. The README explicitly warns not to confuse these test scores with the validation-set trajectories or the historical WCHW 'After' value in a comparison table. That warning suggests the project has faced confusion about which numbers are final. The cost figures are attractive: under $5 per task for the full search and under $0.001 per problem for inference. However, these costs depend on the specific LLM pricing for Claude and Qwen models, which can change. The README does not provide a breakdown of how those costs were calculated, so you should verify with current API rates.

## Limitations and Failure Modes

The project has several genuine limitations. First, the optimizer relies on a specific LLM (Claude-Opus-4.5) for the MCTS search, which means the optimization process is not deterministic and could produce different workflows across runs. The README mentions that the MCTS enhancements reduce noise, but they do not eliminate it. Second, the benchmark suite is narrow: all tasks are wireless communication problems. The README claims the system outperforms a general-purpose workflow optimizer, but that comparison is only on WirelessBench. For non-wireless tasks, the domain-specific tools (ray-tracing, Kalman filters, telecom calculators) would be irrelevant, and the search might not transfer. Third, the held-out test scores are only reported as aggregate numbers; the README does not provide per-problem error analysis or confidence intervals. Fourth, the project has only one release (v1.0.0) and a recent last push, so maintenance is unproven. The README does not mention any known failure cases, but a likely failure mode is that the MCTS search could overfit to the validation split if the experience replay buffer is not carefully managed. The 3-class replay is designed to mitigate noise, but it is not a guarantee.

## Alternatives and the Difference in Approach

A direct alternative is manual agent design, where a human iteratively crafts prompts and tool chains through multi-round dialogue with an LLM. The README explicitly contrasts this with WirelessAgent++: on the left, a human expert designs a fixed workflow; on the right, the optimizer LLM searches over structures automatically. The difference is that manual design is labor-intensive and may miss non-obvious tool combinations, while MCTS explores a broader space but at the cost of requiring an optimizer LLM and compute. Another alternative is a general-purpose workflow optimizer, which the README claims WirelessAgent++ beats by 11.1 points. The key difference is domain adaptation: WirelessAgent includes domain-specific tools and a benchmark suite, whereas a general optimizer would treat wireless tasks as generic text problems. If you already have a well-tuned manual workflow, the MCTS search might not beat it by much, especially if your task is simple. The README's own data shows that the largest gains come from tool discovery, so if your task does not have tools to discover, the benefit shrinks.

## Conclusion

Adopt WirelessAgent-R2 if you work on wireless communication tasks where accuracy gains of 10 to 30 percentage points justify the search cost (under $5 per task) and you can accept the dependence on specific LLMs (Claude-Opus-4.5 for optimization, Qwen-turbo-latest for execution). Do not adopt it if your tasks are outside wireless domains or if you need deterministic, LLM-free operation from the start; the compiled CodeLevel operators are a step in that direction but the optimizer itself is LLM-driven. Before adopting, verify that the reported held-out test scores (78.37% WCHW, 90.95% WCNS, 97.07% WCMSA) reproduce on your own split, confirm the exact cost figures with your API pricing, and check whether the MCTS enhancements (penalized Boltzmann, maturity-aware critic, 3-class replay) are configurable in the codebase. The repository is MIT-licensed, so integration is permissive, but the maintenance cost is unknown given the single release and recent last push.

## FAQ

### What does WirelessAgent-R2 actually optimize?

The structure of the agent workflow itself. An Optimizer LLM, named as Claude-Opus-4.5, searches over compositions of five operators with MCTS, and the discovered workflow is executed by a separate Executor LLM, named as Qwen-turbo-latest, on the benchmark.

### How does WirelessAgent-R2 remove the LLM from tool calling?

MCTS discovers that a ReAct `ToolAgent` loop repeatedly resolves to the same deterministic tool, and then compiles that loop into a `CodeLevel` pipeline. On WCNS the compiled `CodeLevelRayTracing` stage added 1.7 points over the tool-calling version while removing the model from the path.

### How large is the WirelessAgent-R2 benchmark suite?

3,392 problems across three benchmarks. WCHW holds 1,392 problems split 348 validation and 1,044 test, WCNS holds 1,000 split 250 and 750, and WCMSA holds 1,000 split 250 and 750.

### What does a full WirelessAgent-R2 optimization run cost?

One WCHW run is reported as 19 rounds costing $4.95 in total search at roughly 63 minutes wall-clock, with per-problem inference under $0.001. The WCHW validation score went from 62.44 percent at round 1 to 81.78 percent at round 14.

## Sources

- [Official README](https://github.com/jwentong/WirelessAgent-R2#readme)
- [Project repository](https://github.com/jwentong/WirelessAgent-R2)
- [Release notes](https://github.com/jwentong/WirelessAgent-R2/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/jwentong-wirelessagent-r2