WirelessAgent-R2

What the project aims to do

WirelessAgent-R2 is a research repository from the authors of the WirelessAgent++ system, which automates the design of LLM based autonomous agents for wireless network tasks. Building these agents by hand is slow and depends on expert intuition, so the project frames agent design as a program search problem and solves it with a domain adapted version of Monte Carlo Tree Search. An optimizer model jointly searches over workflow structure and tool calling strategies, while a separate executor model runs the resulting workflow against a benchmark. The README reports that the discovered workflows outperform hand crafted baselines by up to 31 percentage points and beat a general purpose workflow optimizer by 11.1 points. The total search cost stays low, with the project citing under five US dollars per task and under one tenth of a cent per problem in inference. The work targets three wireless problem families: knowledge reasoning about communication homework, network slicing with code and tool use, and mobile service assurance that chains prediction steps. The executor model named in the configuration is Qwen Turbo, while the optimizer uses a Claude model, and the arXiv paper linked from the README describes the method in full. The repository therefore serves both as a runnable artifact and as a companion to the published research, so a reader can reproduce the numbers rather than trust a claim. The combination of a cheap search and a clear benchmark is what makes the work useful to other researchers who want to test agent design on a domain with hard, measurable answers. The combination of a cheap search and a clear benchmark is what makes the work useful to other researchers who want to test agent design on a domain with hard, measurable answers rather than vague impressions.

How the optimizer searches for workflows

The optimizer builds workflows from a small set of operators rather than editing raw prompts. Custom invokes an LLM with an instruction, ToolAgent runs a ReAct style agent that interleaves reasoning with tool calls, and CodeLevel runs deterministic tool code with no LLM in the loop, which keeps variance near zero and cost minimal. ScEnsemble votes across candidate answers for self consistency, and AnswerGenerate produces the final structured reply. Three domain aware changes sit on top of standard Monte Carlo Tree Search. A penalized Boltzmann selection stops the search from getting stuck in weak subtrees, a maturity aware critic pre screens proposed mutations with a lightweight model before the expensive evaluation, and a three class experience replay labels outcomes as success, neutral, or failure instead of a binary signal. The README notes a concrete finding from the search: the optimizer learns to compile ToolAgent workflows into CodeLevel pipelines, removing the LLM from the tool calling path while keeping accuracy. The project also lists its domain tools, including ray tracing predictors and Kalman filters, which the agents can call through the ToolAgent operator during a run. This matters because the compiled pipeline runs faster and cheaper at inference time, so the search pays a one time cost to find a workflow that then serves many queries at low expense. This matters because the compiled pipeline runs faster and cheaper at inference time, so the search pays a one time cost to find a workflow that then serves many queries at low expense.

Benchmarks and running the code

WirelessBench provides the evaluation ground, with 3,392 problems split across WCHW, WCNS, and WCMSA. The data pipeline seeds problems from wireless textbooks such as Goldsmith and Molisch and from 3GPP and IEEE standards, then cleans them with a multi LLM funnel and validates every item with graduate students. Reported held out test scores are 78.37 percent for WCHW, 90.95 percent for WCNS, and 97.07 percent for WCMSA. To run the code, a contributor clones the repository, creates a conda environment with Python 3.9, and installs the requirements. API keys for the optimizer and executor models go into a copied config file, and datasets download with a provided script. The run command accepts a dataset name and a maximum round count, so a user can reproduce the WCHW search with nineteen rounds or try general NLP sets such as MATH and HotpotQA. The repository also mirrors the AFlow framework it builds upon, and the project structure separates benchmarks, scripts, and operator definitions so the search logic stays easy to navigate and modify. The README documents the command line arguments in a table, so a new user can set sample count, validation rounds, and convergence checks without reading the source. The README documents the command line arguments in a table, so a new user can set sample count, validation rounds, and convergence checks without reading the source code first.

Editorial conclusion

WirelessAgent-R2 is released under the MIT license and the README pairs the code with an arXiv paper describing the WirelessAgent++ method.