LOS ANGELES, Calif. November 12, 2026. Artificial intelligence is becoming remarkably good at writing code. But scientific discovery requires much more than code that works on a laptop.

Researchers routinely run complex analyses across clusters, research clouds, supercomputers, and distributed computing systems. Data must move between machines; specialized software environments must be created; hundreds of jobs may need to be coordinated; and failures often have to be diagnosed across multiple layers of infrastructure.

PegasusAI Studio, a new open-source platform developed through a National Science Foundation (NSF)-funded collaboration among the University of Southern California (USC), the Renaissance Computing Institute (RENCI), the University of Massachusetts Amherst (UMass), and the University of Tennessee, Knoxville (UTK), is designed to help bridge that gap.

PegasusAI Studio gives AI agents the knowledge and tools they need to help scientists build, run, monitor, and debug computational workflows on real research cyberinfrastructure.

Beginning November 12, 2026, researchers will be able to use PegasusAI Studio through a publicly available science gateway at https://studio.pegasus-ai.org. Sign-in uses CILogon with institutional credentials or another supported provider. Accounts are usually approved within one business day. Once approved, a workspace is set up automatically, and a guided tour appears on first login. The platform also allows institutions to deploy Studio in their own computing environments.

The work is described in the paper “PegasusAI Studio: Legible Cyberinfrastructure and Accountable AI Agents for Scientific Workflows.”

“AI agents are becoming very capable programmers, but scientific computing is much more than programming,” said Ewa Deelman, principal investigator at USC’s Information Sciences Institute and a research professor at the USC Viterbi School of Engineering and USC Stevens School of Computing and AI’s Thomas Lord Department of Computer Science. “Scientists work across workflow systems, schedulers, software environments, data repositories, clouds, and HPC systems. PegasusAI Studio is about connecting AI to that existing infrastructure rather than asking AI to replace it.”

From code generation to scientific execution

(Image/Courtesy of Ewa Deelman)

PegasusAI Studio builds on the Pegasus Workflow Management System and the HTCondor Software Suite reflecting a long-standing collaboration in workload management and distributed scientific computing between the Pegasus team and the Center for High Throughput Computing led by Miron Livny, John P. Morgridge professor of computer science at the University of Wisconsin-Madison.  Instead of recreating scheduling, data movement, failure recovery, provenance, and distributed execution inside the AI layer, Studio delegates those responsibilities to the underlying cyberinfrastructure. Its role is different: to make that infrastructure understandable and usable by AI agents.

Studio gives agents up-to-date information about the workflow system and computing environment, access to tools for managing and running workflows, and structured details about jobs that are running or have failed.

That becomes particularly important when something goes wrong. Rather than asking an AI agent to infer a failure from terminal output, Studio can provide information already recorded by the workflow and scheduling systems, including job states, scheduler holds, logs, data-staging information, and diagnostic reports. The AI agent can then help diagnose the problem, repair the workflow or its software environment, and resubmit it.

“AI agents become much more useful for science when they have access to the context needed to understand what is actually happening during an execution,” said Komal Thareja, an AI expert at RENCI. “PegasusAI Studio gives agents structured information about the workflow and its computing environment, allowing them to reason about failures and help scientists move from identifying a problem to fixing it.”

Tested on real scientific workflows

The researchers evaluated Studio using applications spanning several areas of NSF-supported computational science.

For weather-radar workflows, Studio helped identify and repair multiple infrastructure and configuration problems while modernizing legacy scientific software.

In a gravitational-wave analysis, an AI-written workflow was turned directly into an interactive Studio application and successfully executed 350 jobs.

An EEG analysis tested Studio with a different AI agent and an open-weight model. After recovering from an execution-environment problem, the workflow completed 127 jobs across data from 24 subjects.

In another test, researchers used data from the Event Horizon Telescope—the international collaboration that produced the first image of a black hole—to reconstruct images of the black hole in galaxy M87. The first two attempts failed because the software containers needed to run the workflow had been built incorrectly. Studio passed details of those failures to an AI debugging agent, which helped identify and fix the problems. The next attempt completed all 55 jobs without a failure and produced the reconstructed images.

AI that remains accountable

PegasusAI Studio is also designed for an emerging challenge: what happens when AI systems begin taking actions on shared computing resources?

Agents in Studio operate under the identity and permissions of the scientist using them. Jobs submitted by an agent remain attributable to that researcher, and higher-risk actions and workflow submissions remain subject to user approval controls.

“AI should not create a parallel path around the identity, security, and resource controls that scientific facilities already depend on,” said Mats Rynge, one of the  PegasusAI Studio’s main architects. “The agent should work through the same trusted systems as the scientist.”

Studio can also work with different AI models and providers, allowing institutions to choose commercial models, research-hosted services, or open models according to their own cost, privacy, and data-governance requirements.

In the near future, the same Studio application will be able to be deployed through JupyterHub on research clouds or through Open OnDemand on HPC systems.

The larger goal is to enable increasingly capable AI systems to participate in NSF-funded science while continuing to rely on the mature cyberinfrastructure, security mechanisms, and workflow systems that research already depends on.

PegasusAI Studio in Education

PegasusAI Studio is also becoming a tool for training the next generation of the scientific workforce. Because Studio lets students build, run, and debug real workflows with an AI agent working alongside them, Michela Taufer, professor of computer science and Sai Swaminathan, assistant professor in computer science at UTK piloted it in an EECS data engineering course in spring 2026 and will fully deploy it there in spring 2027. Together with UTK’s Hack4Impact chapter, they are also using Studio in a series of hackathons where students build soil-moisture workflows and see first-hand how AI is becoming a peer programmer.

“With Studio, students can see exactly what the AI agent is doing and why, which turns it into a true peer programmer they can learn from and question,” said Taufer. “How people and AI agents share control is fundamentally a human-computer interaction question,” said Swaminathan. “Studio lets us study how students build trust in an AI collaborator, when they step in, and how the interface can make the agent’s reasoning and actions easier to understand.”

PegasusAI Studio is open source. The project is supported by the U.S. National Science Foundation under Award #2513101 and is a collaboration among USC, RENCI, UMass, and UTK.

To contact the PegasusAI team, please email [email protected]

Published on November 12th, 2026

Last updated on November 12th, 2026