Paper Introduces TSFLab Workbench To Cut Agent Research Costs

Researchers from a group spanning several US universities have published a paper proposing TSFLab, a workbench built to make coding agents useful for time series forecasting research without the cost and upkeep problems that come with pointing an agent at a research task directly. The paper names Claude Code and Codex as examples of existing coding agents that are suboptimal for this work when applied without modification, citing high running costs, poor scaling to large projects such as full benchmarking suites, and no built-in way to keep the resulting code maintained.

The authors’ starting problem is a familiar one in forecasting research: many published papers do not release working code, so anyone who wants to reproduce or build on prior results has to reimplement it, and independent reimplementations tend to drift apart in their experimental settings. The paper reports that only about 20% of arXiv time series forecasting papers carried a code link during the period it surveyed, which leaves most published results without a shared, checkable implementation.

TSFLab is described as agent-agnostic, meaning it is designed to equip a range of coding agents, rather than one vendor’s tool, with shared resources and tooling for reproduction, benchmarking and new method development, and to maintain those resources on an ongoing basis rather than leaving them to rot once a project ends. The authors report testing it across three settings: reproducing published results, running benchmarks across baselines, and a pilot study of agents developing new forecasting methods.

The numbers

• Only about 20% of arXiv time series forecasting papers in the surveyed period included a code link, which the authors cite as the gap TSFLab is meant to close.
• The first version of the paper, posted 29 October 2026, was recorded by arXiv at 1,137 KB.
• The authors report improvements to task completion and reductions in both fresh tokens spent and development time across their reproduction, benchmarking and method development experiments, though the exact ranges are specific to their chosen baselines and settings.

The paper sits inside a broader pattern of research groups treating commercial coding agents as infrastructure to be wrapped and managed rather than used off the shelf. For a field where most published code either does not exist or decays once its authors move on, a workbench that promises both lower running cost and continued upkeep changes the calculation for anyone deciding whether to reimplement a paper by hand or send an agent after it instead.