Document Type
Lecture
Abstract
Agentic AI offers new opportunities to accelerate scientific discovery for time series by enabling autonomous systems to inspect data, formulate hypotheses, compose experiments, and refine models based on empirical evidence. However, building agents that make sound modeling decisions requires more than connecting a language model to forecasting tools. Key challenges include exposing large model catalogues without overwhelming the agent, grounding parameter choices in data characteristics, supporting iterative experimentation, and evaluating the quality of the reasoning behind each decision.
This hands-on lab introduces participants to agentic scientific discovery for time series using the Model Context Protocol (MCP). Through five interactive sprints, participants will build and run a discovery workflow that progresses from data inspection and evidence-based parameter selection to model composition, diagnostic-driven refinement, benchmarking, and evolutionary search. The lab emphasizes three core design principles: exposing model components as discoverable data, providing evidence while leaving modeling decisions to the agent, and evaluating the decision process alongside predictive performance.
Participants will begin by deploying an MCP server and connecting an agent to a catalogue of classical estimators, time series foundation models, anomaly detectors, transformers, and feature extractors. They will then examine seasonality, stationarity, channel relationships, and missingness to justify choices such as lookback and context length. Building on this evidence, participants will compose executable modeling recipes, inspect individual ensemble members’ performance, repair underperforming combinations, and track the lineage of successive experiments.
The lab also introduces systematic evaluation and search. Participants will benchmark recipes across multiple time series configurations using a GIFT-Eval-style protocol and examine a hidden reasoning rubric that assesses whether modeling choices are supported by the available evidence. A final sprint extends the experimentation loop into an AlphaEvolve-style search process, where agents propose candidate recipes and contribute to a shared archive of evaluated solutions. Throughout the session, participants will examine both successful refinements and cases where added complexity fails to improve on a strong baseline.
The lab uses TSM-MCP-Kit, an open-source MCP server distributed within the AssetOpsBench repository, with sktime as its execution substrate. The Docker-based environment includes curated industrial telemetry, a component catalogue, and a reference agent, providing a practical foundation for reproducible experiments. Participants will leave with reusable tools, modeling recipes, and experiment histories to support further exploration.
Publication Info
Preprint version 41st AAAI Conference on Artificial Intelligence (AAAI-27), ed. AAAI, Spring 2027.
APA Citation
Patel, D., Feng, M., Shyalika, C., Lin, S., & Abdullah, H. (2027). A hands-on lab on agentic scientific discovery for time series with the model context protocol [Preprint]. 41st AAAI Conference on Artificial Intelligence (AAAI-27).
Rights
Copyright © 2027, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.