Skip to main content ->
Ai2

Agent Tools

Asta resources includes the first production-grade set of tools that let agents perform scientific research tasks, with an evaluation mode that enables fair, reproducible evaluation.

Asta Scientific Corpus Tool

A production-quality toolset for searching and navigating a large-scale scientific literature graph containing 225M+ papers, 80M+ authors, 550M+ paper-authorship edges, and 2.4B+ citation edges. This includes a new first-of-its-kind snippet search endpoint, indexing 12M+ full-text publications, totaling 285M+ passages. The literature graph may be accessed via an MCP interface for easy integration with any agent. Use the integration in AstaBench (and no other literature graph tools) for reproducible, fair comparison that limits results to papers dated prior to a benchmark-specified date; submissions that do so are marked “Standard Tools” on the AstaBench leaderboards.

Use in AstaBench
Use outside of AstaBench

ComputationalNotebook

A Jupyter computational notebook tool—the first such tool to enable reproducible evaluation and broad agent compatibility. Use this tool (and no other execution environments) for fair comparison when completing the SUPER experiment reproduction task (SUPER analyzes notebook usage); such submissions are marked “Standard Tools” on the AstaBench leaderboards.

Use in AstaBench

WritingEditors (Experimental)

Editors for composing literature analysis reports and tables, with support for citing academic sources. Use these stateful tools to allow agents to incrementally compose analyses during the research process, just as humans do.

Use in AstaBench


Agents Suite

The most comprehensive suite of AI agents, with high-performing Asta science agents, as well as science and general agent baselines. These agents provide the best foundation for comprehensive agent measurement and future development.

Baseline agents

A large collection of agents—both general and science-optimized—including well-known open-source baselines in the literature and connectors to closed products.

Access the agents code

Asta science agents

Open-source scientific research agents, with demonstrated high performance on AstaBench science tasks.

Included in the Asta science agents are the open source agents that power Asta—Find papers (PF), Summarize literature (SQA), and Analyze data (DV). Each is optimized for particular scientific literature review and analysis tasks, including discovering relevant research, summarizing studies, and generating hypotheses and statistical tests.

Access the agents code
Access a growing set of deployed agents


Other Resources

Additional resources for developing scientific AI agents.

Open Language Models

Open language models that have been post-trained for science, which can power open AI agents that excel at science tasks, such as those in our agents suite.

Access the models

Get started on your own agents

Build your own agent or model

Use or fork a general or science-optimized agent in our agents suite.

Evaluate your own agent or model

Rigorously evaluate agents and models on science research tasks using AstaBench.