TSFLab Workbench Raises Agent Task Completion Rates by 20–163% in Time Series Forecasting

Researchers from the D2I group have introduced TSFLab, a "workbench" for AI agents that work on time series forecasting (TSF). It is infrastructure that sits outside the agent itself: it changes neither the model nor its harness, and it plugs into any compatible coding agent. In paper-reproduction and benchmarking experiments, the workbench raised the task completion rate by 20–163% and cut fresh-token consumption per completed task by 11–66% compared with baseline configurations.

Why agents struggle with TSF

The authors start with a problem in the field itself. Forecasting papers keep multiplying, but code is rarely released: of 1707 arXiv papers since 2020, roughly one in five links to code. Researchers, especially newcomers, have to re-implement baseline methods by hand and assemble evaluation pipelines. Independent implementations can diverge in experimental settings, and they are hard to maintain.

Libraries of ready-made methods are not keeping pace with the flow of papers either. According to the paper, one of them grew from 18 to 41 methods, even though more than 1200 new papers have appeared since 2024. In April 2026, its authors announced that they no longer maintain the project because they lack maintainers.

Applying coding agents directly runs into two problems. The first is slowly updated knowledge: a model knows only what was in its training data, and repeated web search is expensive in tokens on large projects. The sources it finds can be incomplete or unreliable, and unverified errors carry over into later work. As an example, the authors cite the official code of one model, which mistakenly dropped the last test batch during evaluation. That affected the settings of many subsequent studies.

The second problem is inconsistent settings. Agents that configure experiments independently rely on different sources. The result is different setups and implementations, results that cannot be compared, and teams that needlessly repeat experiments.

What the workbench consists of

The main difference from earlier work is that the workbench sits outside the agent. Existing TSF agents develop their own harness and gain nothing automatically when newer agents appear. The workbench, by contrast, can be attached to any compatible agent and benefit from that agent's growing capabilities. A reproducer agent adds verified implementations to the workbench, a benchmarker agent uses them, and a third agent takes the ready-made benchmarks to develop new methods.

TSFLab has four modules. TSF-Domain Guide introduces the agent to standard procedures, evaluation protocols and known pitfalls: introductory instructions are supplemented by 17 task-specific skills. For example, when a new dataset is added, the skill asks the agent to identify the time fields, the target variables and the frequency, and then to check whether the forecast is evaluated on a single target or on all variables. This keeps metrics consistent across methods.

Knowledge Structurer breaks existing models down into reusable blocks and assembles methods, blocks and datasets into a single structure. Each element has a "knowledge card" that is read in layers: summary, card, source. Summaries are for selection, cards describe capabilities, parameters, interfaces and limitations, and the source layer holds papers and reference code. The agent takes as much detail as a decision requires and does not read all the code and data. The workbench integrates 323 models and 93 datasets, and shared components are extracted into 71 blocks; 263 methods use at least one of them.

Experiment Workflower runs the experiments the agent specifies through a single runner. It validates configurations and reports results under unified protocols. Recursive Improver handles self-maintenance: it regularly checks each method against its paper and official code, and merges verified methods, blocks, datasets and findings into the shared workbench. It updates independently of the release cycles of the agents themselves.

What the experiments showed

The experiments covered reproduction, benchmarking and a pilot study of method development. In the first two tasks, the workbench raised the task completion rate by 20–163% and cut fresh-token consumption per completed task by 11–66% across all baseline configurations. In the method-development pilot, agents with the workbench spent 32.8–73.6% fewer fresh tokens and 9.2–68.8% less time than in the other modes.

The authors stress that a correct result here means a task done exactly as required, and that cost is counted per completed task. For teams building research agents, the takeaway is practical: a shared, verifiable knowledge layer and a unified experiment runner save tokens and make the results of different agents comparable. The code is published in the D2I-Group/TSFLab repository on GitHub.