Agents in Research
« Agents in Research

#15 — NSF AI funding FY26, Agentic AI data problems, Robotics LLM discovery

June 22, 2026
The National Science Foundation has specified its artificial intelligence investment priorities for fiscal year 2026, covering research, education, and infrastructure. Concurrently, many agentic AI projects are stalling due to pervasive data-related problems. In scientific applications, large language models are being integrated with robotic platforms to create autonomous discovery systems. The use of generative AI in academic grant applications is also increasing competition and changing the nature of proposal review.

The National Science Foundation (NSF) detailed its AI investment priorities for fiscal year 2026. These include fundamental and applied AI research, AI education and workforce development, and access to computing infrastructure. The NSF aims to advance AI, support U.S. leadership, and apply AI for discovery, innovation, and economic growth.

A report indicates many agentic AI projects are stalling or being abandoned due to widespread data problems. Identified obstacles include skill gaps, large language model reliability issues, insufficient data infrastructure, poor data quality, and governance challenges. These issues also create security and integrity risks. One mitigation approach involves integrating data processing and governance earlier in the development cycle.

Large language models (LLMs) now integrate with robotic platforms to form autonomous scientific discovery systems. These systems allow AI agents to automate hypothesis generation, experiment planning, and iterative validation, facilitating closed-loop discovery in molecular science.

Generative AI tools are driving a rise in academic grant applications, leading to a more competitive funding landscape by 2026. This increased volume presents challenges for funding bodies. While AI tools improve the initial quality of research proposals, this also increases the review workload, requiring proposals to show deep domain knowledge and distinct methods to stand out.

AI tools are expected to reduce the time spent on grant applications by 2026. They streamline the production process for research administration without replacing core expertise. These tools can generate initial drafts, restructure text, adapt language, and create supplementary content like budget narratives for grant applications. However, AI tools do not replace human strategic functions in grant applications, such as understanding funder priorities, presenting evidence of community need, or applying strategic judgment. These elements remain human responsibilities in research administration.

NASA tested its ERNEST prototype rover, which covered 16 miles in the Colorado Desert with minimal human input. This four-wheeled rover enhances robotic autonomy and navigation for difficult terrains, intended for lunar and Mars missions. The Materials Acceleration Platform for Electrochemistry (MAP-E) is an autonomous system for high-throughput electrochemical experiments, such as corrosion testing. It combines robotic liquid handling and multi-channel control to obtain corrosion data without human involvement. AutoScientists identified a method that improved ACE2-Spike binding prediction by 12.5% in Spearman correlation over existing models, demonstrating capability in protein fitness prediction using ProteinGym.

Sources

  1. AI funding FY 2026 - NSF
    NSF's AI investments for FY26 will concentrate on three interconnected areas: fundamental and translational AI research, education and workforce development, and access to data and advanced computing research infrastructure and testing platforms. The goal is to advance AI, cement U.S. leadership, and strengthen the ability to harness AI for discovery, innovation, and economic growth. This includes supporting foundational research in various AI subfields, fostering use-inspired and translational research through initiatives like the National AI Research Institutes program, and developing the…
  2. Most agentic AI projects in production have stalled over data problems - Help Net Security
    Many agentic AI projects in production, despite a rise in adoption, are stalling or being abandoned due to pervasive data problems. Key obstacles include skills gaps, LLM reliability, data infrastructure and quality issues, and governance challenges, which collectively present significant security and integrity concerns. The report suggests "shifting left" by moving data processing and governance closer to the data source to address these issues.
  3. Transforming Molecular Science With Large Language Models: From Molecule Understanding to Autonomous Scientific Discovery
    Large language models (LLMs) are driving a paradigm shift in molecular science, transitioning from molecule understanding to autonomous scientific discovery. By learning from multimodal molecular representations through advanced training techniques, LLMs address traditional limitations in molecular research, including expert dependency and limited experimental scalability. We analyze strategies for cross-modal alignment and domain adaptation that enable LLMs to interpret complex molecular semantics. This deep semantic understanding translates into broad versatility across key downstream…
  4. Surviving the 2026 Grant Writing Surge: When AI Is Not Enough ...
    The year 2026 presents a highly competitive environment for academic funding due to an unprecedented surge in grant applications, largely fueled by generative AI tools. While AI has elevated the baseline quality of proposals, it has created a "burden of choice" for reviewers, necessitating that research stands out through profound domain expertise and unique methodological approaches. To succeed, researchers must move beyond AI drafting, focusing on rigorous, deeply collaborative project management and centralizing their grant strategy with tools like ResearchDock.
  5. AI Tools for Grant Writing and Funding Applications in 2026: What's ...
    AI tools in 2026 are not replacing the expertise and strategy behind successful grant applications, but they significantly reduce the time needed to produce them. They excel at first draft generation, content restructuring, adapting language, and generating supporting content like budget narratives. However, AI cannot replace deep knowledge of funder priorities, authentic evidence of community need, or strategic judgment.
  6. NASA Testing Advanced Capabilities for Moon, Mars Rovers
    A prototype rover built with a new design for tackling rugged terrain is helping teams refine capabilities that could one day be used on future lunar and Red Planet missions. On a bleak stretch of the Colorado Desert in Southern California, a compact four-wheeled rover recently trundled about 16 miles (26 kilometers) with minimal intervention from the team of engineers trailing it. Called ERNEST (Exploration Rover for Navigating Extreme Sloped Terrain), this prototype is being used by NASA to advance both robotic autonomy and the ability to traverse challenging landscapes.
  7. Materials Acceleration Platform for Electrochemistry: a Platform for Autonomous Electrochemistry
    Abstract Corrosion testing is slow, labor-intensive, and sensitive to operator technique, limiting the generation of large, high-quality datasets for data-driven materials discovery. The Materials Acceleration Platform for Electrochemistry (MAP-E) is an autonomous, high-throughput system, capable of performing parallel electrochemical experiments. It integrates robotic liquid handling and sample transfer with a multi-channel potentiostatic control to extract corrosion metrics without human intervention. Validation against an ASTM G61-analog benchmark demonstrates good reproducibility, with a…
  8. AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation
    Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision. AI agents can automate parts of this process, but existing approaches typically follow a single research trajectory or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration, adapt as experimental evidence changes, or preserve knowledge of failed directions over long-running experiments. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation.…

Also this week

Full transcript
Where is the US government directing its artificial intelligence funding for the next fiscal year? The National Science Foundation has outlined its priorities, and that's our lead story today on Research Agent. We'll start there. The National Science Foundation has laid out its AI investment priorities for fiscal year 2026, and they're focused on three core areas. Right, it’s fundamental research, education and workforce development, and access to computing infrastructure. The stated goal is to push AI forward and maintain U.S. leadership. But there's a disconnect between that high-level planning and what's happening on the ground. While the NSF is looking to the future, many current agentic AI projects are reportedly stalling out. And it's not a mystery why. The primary cause seems to be pervasive data problems. Things like a lack of skills, unreliable models, and just plain insufficient data infrastructure and quality. So you have these big ambitions running into very practical, foundational roadblocks. It's not just data, either; the report mentions governance challenges and security risks as well. Which is why some are suggesting a 'shift left' approach—integrating data processing and governance closer to the source, to catch these problems earlier. But even with these hurdles, some researchers are pushing ahead to build more capable systems. How? What are they building? A recent paper details integrating large language models directly with robotic platforms. The idea is to create a closed-loop system for scientific discovery. Meaning the AI does everything? From hypothesis to experiment to validation, all automated? Exactly. It's being applied to create autonomous workflows in molecular science, to get past the limits of scale and reliance on human experts. And we’re seeing specific examples pop up. Like what? There’s a system called MAP-E, the Materials Acceleration Platform for Electrochemistry. It’s a high-throughput system for autonomous corrosion testing, combining robotic liquid handling and multi-channel control to get metrics without a human in the loop. And it was successful? It worked as a proof of concept? It did. It showed good reproducibility against a standard benchmark and was able to autonomously construct stability diagrams for stainless steel. And that's not the only domain. Right, NASA is doing something similar with robotics for planetary exploration. Yes, with its ERNEST prototype rover. They tested it in the Colorado Desert, where it drove 16 miles with minimal human direction. The goal is to refine autonomous navigation for difficult terrains on the Moon and Mars. And speaking of tangible results, the AutoScientists agent team we mentioned earlier this month actually made a discovery. They did. The system found a method that improves the prediction of ACE2-Spike binding by 12.5% over the previous state-of-the-art model. So these systems are producing novel scientific outcomes. So the AI is changing the science, but it’s also changing the *administration* of science. You mean how research gets funded. Generative AI tools are reportedly causing a huge increase in academic grant applications. Which is creating a highly competitive environment. The tools are elevating the baseline quality of proposals, creating what sources call a 'burden of choice' for reviewers. So to stand out now, you need more than just a well-written proposal. You need deeper domain expertise and unique methodologies that the AI can't just generate for you. The tools can help with drafts and supporting materials, but this confirms a point we've made before: AI can't replace human expertise. Strategic thinking, understanding what a funder wants—those remain human tasks. And there are a few other developments happening in this space. Anthropic is out there urging its rivals to pursue AI arms control. At the same time, researchers are using video games and virtual environments to do more research on AI agents. And on the business side, Dun & Bradstreet just introduced agentic AI into its risk analytics platform. There’s even a proposal to use an AI platform with drone swarms to address audit governance in Saudi Arabia. Which fits with another study that introduced a framework for autonomous regulatory monitoring. So it’s touching everything from basic science to corporate governance. That's all for this update. We will track these developments as they continue. Until next time, on Research Agent.