Agents in Research
« Agents in Research

#4 — The Rise of AI Agents in Scientific Discovery

June 11, 2026
Agentic AI systems are being developed to accelerate scientific discovery. New platforms and frameworks use multi-agent orchestration and large language models to automate tasks from bioinformatics to proposing new experiments. Studies indicate these systems can reduce task completion times and the number of experiments required to find optimal results. However, their adoption introduces challenges regarding reproducibility and ethics, and underscores the crucial role of human experts for methodological guidance and synthesizing insights.

Sources

  1. PromptBio Launches Agentic AI Platform to Accelerate Scientific ...
    New multi-agent AI platform combines conversational research, bioinformatics, and drug discovery workflows to transform scientific questions into actionable insights within a unified environment. The PromptBio Platform combines conversational AI with multi-agent orchestration to accelerate discovery from hypotheses to insights.
  2. AI Autonomous Science Discovery Could Redefine How ...
    AI Autonomous Science Discovery is rapidly transforming scientific research, moving from AI as a calculation tool to an autonomous agent capable of proposing hypotheses, designing experiments, and writing manuscripts without human intervention. This shift, exemplified by systems like Robin and Google DeepMind's Co-Scientist, dramatically reduces discovery time and uncovers cross-disciplinary connections humans often miss. However, it introduces challenges such as reproducibility issues due to stochastic reasoning paths and ethical questions regarding inventorship, safety, and the evolving…
  3. AI Agents Reshape Knowledge Work, Increasing Autonomy and ...
    An arXiv paper quantifies how agentic systems change knowledge work, using production data from Perplexity's Search and Computer products. The paper indicates that Computer performs significantly more autonomous work per user session and has lower dissatisfaction rates compared to Search. It also reports substantial reductions in task completion time and cost, with users shifting towards higher-order verification and extension tasks.
  4. Evaluating agentic AI for biological discovery in autonomous and copilot settings
    Advances in large language models (LLMs)-based artificial intelligence (AI) agents have improved their ability to execute structured analytical workflows, including standard bioinformatic pipelines for biological discovery. However, computational biology rarely consists of deterministic pipeline execution alone. Biological datasets are heterogeneous and noisy, and meaningful discovery often requires open-ended hypothesis generation and iterative reasoning over multimodal evidence. These challenges are particularly evident in multi-omic studies, where paired molecular modalities and…
  5. Beyond AI as Assistants: Toward Autonomous Discovery in Cosmology
    Recent advances in artificial intelligence (AI) agents are pushing AI beyond tools toward autonomous scientific discovery. We discuss two complementary agentic systems for cosmology: \texttt{CMBEvolve}, which targets tasks with explicit quantitative objectives through LLM-guided code evolution and tree search, and \texttt{CosmoEvolve}, which targets open-ended scientific workflows through a virtual multi-agent research laboratory. As preliminary demonstrations, we apply \texttt{CMBEvolve} to out-of-distribution detection in weak-lensing maps, where it iteratively improves the benchmark score…
  6. Best AI Tools for Universities & Research Institutions 2026 - Chitika
    This guide evaluates the ten best AI tools available to universities and research institutions in 2026, reviewing each across criteria like accuracy, citation support, ease of use, and security, with the goal of aiding confident, well-informed AI adoption decisions. It highlights tools such as CustomGPT.ai for research knowledge management, ChatGPT for general writing, and Claude for nuanced document analysis. The article concludes with a buyer checklist, an ROI framework, a risk overview, and best practices for implementation, emphasizing that most AI risks are mitigable through thoughtful…
  7. AI News Roundup: May 26 – June 07, 2026
    This AI News Roundup covers key developments from May 26 to June 07, 2026, including OpenAI's plans for a ChatGPT superapp overhaul, Anthropic's call for a coordinated AI pause plan, and SpaceX's AI compute deal with Google. Other significant news includes Japan's warning about becoming an "AI colony," South Korea's push for AI profit-sharing, and Microsoft's introduction of new in-house AI models and a personal work agent. The roundup also highlights developments in AI regulation, investment, and ethical considerations across various global entities.
  8. Reka and Moonvalley Join Forces to Advance Models and ...
    Reka, an AI research lab, has partnered with Moonvalley, integrating their team of AI researchers and engineers to accelerate the development of models and infrastructure for physical AI. This merger brings experts from DeepMind, Meta, and Google to Reka, with Mateusz Malinowski and Mikołaj Bińkowski joining Reka's research leadership. The combined team will focus on the World Language Action Model (WLAM), an omni-model trained on physical world data for realistic simulation and planning in real-world scenarios.
  9. Training-free active learning framework in materials science with large language models
    Active learning (AL) accelerates scientific discovery by prioritizing the most informative experiments, but traditional machine learning (ML) models used in AL suffer from cold-start limitations and domain-specific feature engineering, restricting their generalizability. Large language models (LLMs) offer a new paradigm by leveraging their pretrained knowledge and universal token-based representations to propose experiments directly from text-based descriptions. Here, we introduce an LLM-based active learning framework (LLM-AL) that operates in an iterative few-shot setting and benchmark it…

Also this week

Full transcript
AI systems are now being designed to operate like autonomous researchers. This shift toward agentic AI is what we are examining on Research Agent. We'll start with how these new platforms work and the challenges they present. [LEAD] PromptBio Platform integrates conversational AI and multi-agent orchestration The PromptBio Platform is a new multi-agent AI system. It combines conversational AI with multi-agent orchestration to accelerate scientific discovery. The platform unifies workflows for conversational research, bioinformatics, and drug discovery, transforming scientific questions into actionable insights. [LEAD] Autonomous science discovery introduces challenges The adoption of autonomous science discovery presents specific challenges. These include issues with research reproducibility due to the stochastic nature of AI reasoning. Ethical considerations regarding inventorship, safety, and the changing role of human scientists are also introduced. [LEAD] An arXiv paper quantified changes in knowledge work due to agentic systems A recent arXiv paper analyzed production data from Perplexity's Search and Computer products to quantify the impact of agentic systems on knowledge work. The study found that Perplexity's Computer, an agentic system, handled more autonomous work per user session and resulted in lower dissatisfaction rates compared to its Search product. It also reported substantial reductions in task completion time and cost, enabling users to reallocate their efforts towards higher-order verification and extension tasks. [LEAD] Researchers developed the Multistep Multimodal Multiomic Agentic (M3A) Framework. The M3A Framework supports large language model-driven reasoning across persistent multimodal data. It also captures agentic reasoning behaviors in autonomous and human-AI copilot settings. This framework facilitates the evaluation of AI systems for biological discovery. [LEAD] Research determined current AI agents effectively explore complex data but require human expertise. An evaluation found that current AI agents competently perform broad, systemic exploration of complex multi-omic data. However, the study concluded that human domain experts are crucial for providing methodological guidance. Experts are also needed for synthesizing biological insights across various analyses. [LEAD] CosmoEvolve demonstrates autonomous ACT DR6 data analysis. CosmoEvolve was demonstrated for autonomous data analysis of ACT DR6 data. It identified non-trivial pair- and scale-dependent behavior. The system produced analysis-grade diagnostics. [NOTABLE] A guide evaluates AI tools for universities and research institutions. The guide assesses ten artificial intelligence tools for use in universities and research institutions in 2026. It reviews each tool based on criteria including accuracy, citation support, ease of use, and security. The guide aims to support informed decisions for AI adoption and highlights specific tools like CustomGPT.ai, ChatGPT, and Claude for various research applications. [NOTABLE] Microsoft introduced new in-house AI models and a personal work agent. Microsoft launched new artificial intelligence models developed internally. Concurrently, the company introduced a personal work agent designed to assist users. These releases expand Microsoft's AI product offerings. [NOTABLE] Reka's combined team will focus on developing the World Language Action Model (WLAM). The newly combined team at Reka will direct its efforts towards developing the World Language Action Model (WLAM). This omni-model is designed to be trained on physical world data. The model aims to enable realistic simulation and planning in real-world scenarios. [NOTABLE] Researchers introduced an LLM-based active learning framework. A new active learning framework, LLM-AL, was developed. This framework utilizes large language models (LLMs) to propose experiments directly from text-based descriptions. It leverages LLMs' pretrained knowledge and operates in an iterative few-shot setting. [NOTABLE (FOLLOW-UP)] LLM-AL outperformed traditional machine learning models in active learning. The LLM-based active learning framework (LLM-AL) was benchmarked against conventional machine learning models. This comparison was performed across four distinct materials science datasets. LLM-AL consistently showed superior performance. [NOTABLE (FOLLOW-UP)] LLM-AL reduced the number of experiments needed for optimal candidates. The LLM-based active learning framework (LLM-AL) significantly decreased the experimental effort required. It reduced the number of experiments needed to reach top-performing candidates by over 70%. This demonstrates its efficiency in scientific discovery. [NOTABLE (FOLLOW-UP)] LLM-AL performs broader and more exploratory searches. The LLM-based active learning framework (LLM-AL) was observed to execute a more expansive and exploratory search strategy. Despite this broader exploration, it still achieved optimal results with fewer iterations compared to traditional methods. [NOTABLE (FOLLOW-UP)] LLM-AL demonstrated consistent performance stability across runs. Researchers examined the stability of LLM-AL given the inherent non-determinism of large language models. Its performance was found to be broadly consistent across multiple runs. This variability aligns with typical observations for traditional machine learning approaches. We'll continue to follow how these systems are being applied in scientific discovery. Until next time, on Research Agent.