#13 — LifeSciBench, NeuralTrust AI security funding, ProteinGym prediction gain

To test an AI's scientific ability, you have to see if it can make judgments from incomplete evidence. That is the standard behind a new benchmark we are examining today on Research Agent. We will start with its use in the life sciences. [LEAD] LifeSciBench provides a benchmark for AI in life science research. LifeSciBench is a benchmark designed to assess AI system performance on complex, real-world life science research tasks. It features 750 tasks across seven workflows and biological domains, developed by 173 scientists. The benchmark evaluates AI's ability to interpret incomplete evidence and make scientific judgments, rather than focusing on fact recall. SOURCE MATERIAL: (Introducing LifeSciBench | OpenAI, 2026-06-17): LifeSciBench is an expert-written, expert-reviewed benchmark designed to assess how well AI systems perform complex, real-world life science research tasks. It comprises 750 tasks across seven workflows and biological domains, developed by 173 scientist contributors with Ph.D.-level training and industry experience. The benchmark focuses on measuring AI's ability to support realistic research, interpreting incomplete evidence, and making scientific judgments rather than just answering fact-recall questions. [LEAD] NeuralTrust secured $20 million in seed funding to enhance its platform for securing AI agents. NeuralTrust, an EU cybersecurity company, raised $20 million in seed funding. This investment is intended to improve its platform designed for securing enterprise AI agents. The funding will support engineering efforts, product integration, and market expansion across Europe. SOURCE MATERIAL: (AI Agents: AI Security for Enterprise AI - Startup Valley, 2026-06-17): NeuralTrust has secured $20 million in seed funding, the largest cybersecurity seed financing raised by an EU company to date, to enhance its platform for securing AI agents in enterprises. This investment will support engineering, deepen product integration, and facilitate expansion across the European market as autonomous AI systems move into production. The platform aims to address the governance gap in AI agent adoption, which is outpacing security measures, by providing tools to monitor, control, and secure AI agents across various enterprise systems. [LEAD (FOLLOW-UP)] AutoScientists improves ACE2-Spike binding prediction by 12.5% on ProteinGym On the ProteinGym fitness prediction tasks, AutoScientists discovered a method that improved ACE2-Spike binding prediction by 12.5% in Spearman correlation over the current state-of-the-art model. This same method, applied across all 217 ProteinGym assays without modification, improved over the prior state of the art by an average of 6.5% in Spearman correlation. SOURCE MATERIAL: (AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation, 2026-05-27): 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. Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, c [LEAD] Researchers developed a stateful ReAct agent using LangGraph to optimize the autoresearch pattern. The autoresearch pattern previously operated in a stateless manner, leading to high token costs during iterative code modification. This work introduced a stateful ReAct agent, built with LangGraph, to reformulate this pattern. The new agent carries experimental history across iterations using typed persistent state and a tool-calling interface. SOURCE MATERIAL: (Remember, Don't Re-read: Stateful ReAct Agents for Token-Efficient Autonomous Experimentation, 2026-06-12): The autoresearch pattern enables autonomous experimentation by having a large language model (LLM) iteratively modify code to optimize a target metric. Its stateless design, however, reconstructs experimental context from scratch at every iteration, incurring $O(n)$ token cost per iteration and $O(n^{2})$ total. This work reformulates the pattern as a stateful ReAct agent using LangGraph, where typed persistent state carries experimental history across iterations via a tool-calling interface. Two benchmarks are evaluated: hyperparameter tuning (15 iterations, small per-iteration observations) and code performance optimization (40 iterations, large per-iteration observations containing full s [LEAD] EurekAgent's code and results were open-sourced. The code and research results for the EurekAgent system were made publicly available. This initiative supports the broader research community. It facilitates further development and replication of autonomous research agent methods. SOURCE MATERIAL: (EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery, 2026-06-11): LLM-based agents have shown increasing potential in automating scientific discovery. Given an optimizable metric and an execution environment, they can propose, validate, and iterate scientific solutions, and have produced results that outperform human-designed approaches. As model capabilities continue to improve, we argue that the bottleneck for autonomous scientific discovery is shifting from prescribing agent workflows to designing agent environments: the resources, constraints, and interfaces that shape agent behavior. We frame this as environment engineering: building environments that amplify productive behaviors, such as open-ended exploration, systematic artifact management, and int [LEAD] AutoMedBench identified "Validate" as the weakest AI agent workflow stage. Analysis of thousands of runs on AutoMedBench revealed that AI agents perform weakest in the "Validate" workflow stage, while the "Setup" stage is the strongest. This suggests that current agents are more effective at preparing research pipelines for execution than at verifying their reliability. The findings come from evaluating agents across a five-stage workflow for medical-AI tasks. SOURCE MATERIAL: (AutoMedBench: Towards Medical AutoResearch with Agentic AI Models, 2026-06-01): Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spann [LEAD] An agentic LLM system outperformed a single LLM system in frailty scoring. A study compared an agentic large language model (LLM) system with a single LLM system for automated frailty scoring in surgical patients. The agentic system demonstrated superior performance in binary frailty classification and generating Risk Assessment Index (RAI) scores. This improvement was particularly notable when using lower-parameter LLMs like Llama 3.1 8b and Qwen 2.5 7b. Automating frailty score calculation from Electronic Health Records (EHR) could support prehabilitation strategies and improve clinical efficiency. SOURCE MATERIAL: (Use of AI agents to assess preoperative frailty in cancer patients, 2026-06-03): Assessment of preoperative frailty is essential for risk stratification and optimization of patients undergoing major abdominal surgery for hepatobiliary (HPB) and gastrointestinal (GI) malignancies. Large language model (LLM)-based agents may facilitate automated frailty scoring, thereby supporting targeted prehabilitation efforts to improve surgical outcomes. We compared the performance of an agentic LLM system with a single LLM system in calculating frailty scores and classifying frailty status. Using real-world preoperative notes documented within 60 days prior to surgery, we evaluated both approaches. The agentic LLM system outperformed the single LLM system in 4 out of 6 models for a b [LEAD (FOLLOW-UP)] A survey proposes five evaluation dimensions for AI research. A survey introduces five key dimensions for evaluating AI research: novelty, validity, impact, reliability, and provenance. These dimensions offer a structured approach to assessing the contributions and quality of AI applications in scientific contexts. They provide criteria for a comprehensive assessment. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] Databricks announced Genie One. Databricks introduced Genie One, an agentic coworker designed to automate and orchestrate tasks for business teams by utilizing various data sources. A central component, Genie Ontology, is a live context layer that continuously learns an organization's operations. This system extracts and updates knowledge from internal and external data, AI tools, and workplace applications, allowing Genie One to provide accurate, enterprise-data-grounded answers and actions. SOURCE MATERIAL: (Databricks Launches Genie One: All-New Agentic Coworker for Every Team, 2026-06-16): Databricks announced Genie One, an agentic coworker designed to automate and orchestrate work for business teams using various data sources. At its core is Genie Ontology, a live context layer that continuously learns an organization's business by extracting and updating knowledge from internal and external data, AI tools, and workplace applications. This allows Genie One to provide accurate answers and actions grounded in governed enterprise data, addressing the "context gap" often seen in early enterprise AI solutions. [NOTABLE] Humans perform better with AI agents complementing personal attributes. Researchers at the MIT Initiative on the Digital Economy found that human performance improves when individuals collaborate with AI agents designed to complement their personal characteristics. These characteristics include personality, gender, and country of origin. This complementarity leads to improved outcomes in human-AI teams. SOURCE MATERIAL: (5 things to consider when working with AI | MIT Sloan, 2026-06-16): To get the most out of working with artificial intelligence, consider these insights from researchers at the MIT Initiative on the Digital Economy: It matters how AI is designed. Humans perform better when teamed with AI agents that complement their personality, gender, and country of origin. AI doesn't always give objective advice. When companies use LLMs to make business decisions, the language used to prompt the AI can have a big impact on how the models respond. Taking time to think about why you're acting on AI recommendations can reduce uncritical reliance on it and boost accuracy without adding significant time to tasks. [NOTABLE (FOLLOW-UP)] AI models do not consistently provide objective advice. A finding from the MIT Initiative on the Digital Economy indicates that artificial intelligence models do not always offer objective recommendations. This observation suggests that AI advice can be influenced by various factors. The lack of consistent objectivity is a consideration when using AI systems. SOURCE MATERIAL: (5 things to consider when working with AI | MIT Sloan, 2026-06-16): To get the most out of working with artificial intelligence, consider these insights from researchers at the MIT Initiative on the Digital Economy: It matters how AI is designed. Humans perform better when teamed with AI agents that complement their personality, gender, and country of origin. AI doesn't always give objective advice. When companies use LLMs to make business decisions, the language used to prompt the AI can have a big impact on how the models respond. Taking time to think about why you're acting on AI recommendations can reduce uncritical reliance on it and boost accuracy without adding significant time to tasks. [NOTABLE] Human reflection on AI recommendations improves accuracy. Researchers at the MIT Initiative on the Digital Economy suggest that consciously considering the reasons behind AI recommendations reduces uncritical reliance on AI systems. This practice enhances accuracy in task completion. It does not substantially extend overall task times. SOURCE MATERIAL: (5 things to consider when working with AI | MIT Sloan, 2026-06-16): To get the most out of working with artificial intelligence, consider these insights from researchers at the MIT Initiative on the Digital Economy: It matters how AI is designed. Humans perform better when teamed with AI agents that complement their personality, gender, and country of origin. AI doesn't always give objective advice. When companies use LLMs to make business decisions, the language used to prompt the AI can have a big impact on how the models respond. Taking time to think about why you're acting on AI recommendations can reduce uncritical reliance on it and boost accuracy without adding significant time to tasks. [NOTABLE] OpenAI introduced Deployment Simulation. Deployment Simulation is a pre-deployment safety method developed by OpenAI. It assesses model behavior by replaying past conversations through a candidate model to identify potential failure modes in realistic contexts. This technique helps estimate the frequency of undesired behavior before release and informs mitigation strategies. It also extends to agentic coding through simulated tool calls for comprehensive risk assessment. SOURCE MATERIAL: (OpenAI's Deployment Simulation Extends Pre-Deployment Risk ..., 2026-06-16): OpenAI has introduced Deployment Simulation, a pre-deployment safety method that assesses model behavior by replaying past conversations through a candidate model to identify potential failure modes in realistic contexts. This technique helps estimate undesired behavior frequency before release and has informed mitigation strategies and deployment decisions. It extends to agentic coding through simulated tool calls, providing a more comprehensive risk assessment than traditional evaluations. [NOTABLE] The stateful ReAct agent reduced token consumption by 90% during hyperparameter tuning. Evaluation of the stateful ReAct agent included a hyperparameter tuning benchmark with 15 iterations. The stateful agent consumed 2,492 tokens, achieving a 90% reduction compared to the 24,465 tokens used by its stateless counterpart. This token reduction did not compromise the optimization quality. SOURCE MATERIAL: (Remember, Don't Re-read: Stateful ReAct Agents for Token-Efficient Autonomous Experimentation, 2026-06-12): The autoresearch pattern enables autonomous experimentation by having a large language model (LLM) iteratively modify code to optimize a target metric. Its stateless design, however, reconstructs experimental context from scratch at every iteration, incurring $O(n)$ token cost per iteration and $O(n^{2})$ total. This work reformulates the pattern as a stateful ReAct agent using LangGraph, where typed persistent state carries experimental history across iterations via a tool-calling interface. Two benchmarks are evaluated: hyperparameter tuning (15 iterations, small per-iteration observations) and code performance optimization (40 iterations, large per-iteration observations containing full s [NOTABLE] The stateful ReAct agent reduced token consumption by 52% during code performance optimization. A second benchmark focused on code performance optimization across 40 iterations. In this scenario, the stateful ReAct agent consumed 627,000 tokens. This represents a 52% decrease from the 1,275,000 tokens required by the stateless agent, while delivering comparable optimization results. SOURCE MATERIAL: (Remember, Don't Re-read: Stateful ReAct Agents for Token-Efficient Autonomous Experimentation, 2026-06-12): The autoresearch pattern enables autonomous experimentation by having a large language model (LLM) iteratively modify code to optimize a target metric. Its stateless design, however, reconstructs experimental context from scratch at every iteration, incurring $O(n)$ token cost per iteration and $O(n^{2})$ total. This work reformulates the pattern as a stateful ReAct agent using LangGraph, where typed persistent state carries experimental history across iterations via a tool-calling interface. Two benchmarks are evaluated: hyperparameter tuning (15 iterations, small per-iteration observations) and code performance optimization (40 iterations, large per-iteration observations containing full s [NOTABLE] A paper introduces the Monroe Grading Agent, an AI-powered desktop application. The Monroe Grading Agent (MGA) is a new AI-powered desktop application designed to automatically grade student semester assignments. It compares student submissions against assignment instructions using TF-IDF vectorization, cosine similarity, and a novel cube-root scaling method. This system provides partial-credit-based grading, aiming to address the heavy workload and consistency issues in higher education grading. SOURCE MATERIAL: (Monroe Grading Agent: AI-Powered Automated Assignment Grading in Higher Education, 2026-05-25): Automated grading has emerged as a valuable approach to address the heavy workload and consistency challenges of manual grading in higher education. Instructors often spend countless hours evaluating student assignments, which can delay feedback and introduce subjective inconsistencies. Prior research highlights the efficiency and objectivity of AI-assisted grading, noting its potential to reduce grading inconsistencies by up to 44% (Yigit et al., 2024) and deliver faster, more scalable feedback (Wang et al., 2022). This paper introduces the Monroe Grading Agent (MGA) – an AI-powered desktop application developed to automatically grade student semester assignments by comparing submissions ag [NOTABLE] The Monroe Grading Agent demonstrates significant improvements in grading efficiency and consistency. An evaluation of the Monroe Grading Agent (MGA) using realistic Monroe University assignment data showed substantial improvements in grading efficiency and scoring consistency. Analysis indicated that MGA drastically reduced grading time while maintaining scores comparable to human evaluators. The system also generates structured feedback aligned with assignment criteria. SOURCE MATERIAL: (Monroe Grading Agent: AI-Powered Automated Assignment Grading in Higher Education, 2026-05-25): Automated grading has emerged as a valuable approach to address the heavy workload and consistency challenges of manual grading in higher education. Instructors often spend countless hours evaluating student assignments, which can delay feedback and introduce subjective inconsistencies. Prior research highlights the efficiency and objectivity of AI-assisted grading, noting its potential to reduce grading inconsistencies by up to 44% (Yigit et al., 2024) and deliver faster, more scalable feedback (Wang et al., 2022). This paper introduces the Monroe Grading Agent (MGA) – an AI-powered desktop application developed to automatically grade student semester assignments by comparing submissions ag [NOTABLE] AutoMedBench showed verification and submission failures dominate AI agent errors. Error analysis from AutoMedBench runs indicated that verification and submission failures are the most common issues for autonomous AI agents, accounting for 37.7% and 38.1% of errors, respectively. Task-understanding errors were rare, at only 0.9%. The study also found that runs with even one error code experienced a 48% reduction in overall score compared to error-free runs. SOURCE MATERIAL: (AutoMedBench: Towards Medical AutoResearch with Agentic AI Models, 2026-06-01): Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spann [NOTABLE] Researchers presented a multi-agent AI system for transcript processing. A multi-agent AI system was developed to automate the processing of diverse high school transcripts for college admissions. This system comprises specialized agents for pattern recognition, semantic analysis, and multimodal document analysis, coordinated by an orchestration agent. Agent-based quality control using GPA extraction ensures reliable collaboration among agents. SOURCE MATERIAL: (A Multi-Agent AI System for Automated High School Transcript Processing: Collaborative Document Analysis at Scale, 2026-06-11): Each year, college admissions offices face an overwhelming challenge: processing millions of high school transcripts, each with unique formats, grading systems, and layouts. This manual process creates operational bottlenecks that delay admissions decisions and consume valuable resources. We present a transformative solution through a multi-agent AI system where specialized agents collaborate to automatically process diverse transcript formats through intelligent coordination and communication. Our multi-agent architecture consists of three specialized agents-a Pattern Recognition Agent for format-specific parsing, a Semantic Analysis Agent for natural language understanding, and a Vision In [NOTABLE] A multi-agent AI system achieved high accuracy and processing speed for transcript analysis. The multi-agent AI system successfully processed 40 real-world high school transcripts from 13 U.S. states. It achieved 96.7% accuracy when compared to expert manual review. The system also maintained a practical processing speed of 45 seconds per transcript. SOURCE MATERIAL: (A Multi-Agent AI System for Automated High School Transcript Processing: Collaborative Document Analysis at Scale, 2026-06-11): Each year, college admissions offices face an overwhelming challenge: processing millions of high school transcripts, each with unique formats, grading systems, and layouts. This manual process creates operational bottlenecks that delay admissions decisions and consume valuable resources. We present a transformative solution through a multi-agent AI system where specialized agents collaborate to automatically process diverse transcript formats through intelligent coordination and communication. Our multi-agent architecture consists of three specialized agents-a Pattern Recognition Agent for format-specific parsing, a Semantic Analysis Agent for natural language understanding, and a Vision In [NOTABLE] Current AI systems in scientific workflows exhibit fragmentation and various limitations. A survey identifies that existing AI systems for scientific workflows are fragmented, showing inconsistencies in autonomy, domain scope, and oversight mechanisms. These systems encounter difficulties with evidence preservation, ensuring reproducibility, rejecting unproductive directions, and tracking provenance. They also struggle with accountable scientific closure. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] A survey organizes AI research systems around five workflow conditions. A survey structures its analysis of AI research systems by categorizing them across five core workflow conditions. These conditions include literature and research grounding, hypothesis formation and planning, experimentation and tool use, feedback, validation, and review, and reporting and knowledge communication. This organization helps analyze how control, evidence, execution, validation, and accountability are managed. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] A review paper suggested methods for trustworthy and scalable Agentic AI systems. The review paper suggested that integrating methods like Agentic Retrieval Augmented Generation (RAG), agent chaining, and formal threat modeling can establish a foundation. These integrations are proposed as crucial for developing the next generation of trustworthy and scalable Agentic AI systems. SOURCE MATERIAL: (A Holistic Review of Agentic AI Frameworks, Applications, and Research Trajectories, 2026-06-15): Abstract Agentic Artificial Intelligence (Agentic AI), as a new generation of intelligent systems, extends beyond mere text or image generation by incorporating components such as multi step reasoning, persistent memory, multi agent interaction, and purposeful tool use. These features enable autonomy and dynamic decision making in open and complex environments. This paper provides an extensive review of the existing literature on Agentic AI. To this end, the conceptual and historical distinctions between Agentic Artificial Intelligence, classical agents, and general-purpose language models are first examined. Subsequently, the proposed architectural framework encompassing perception, role ad [NOTABLE] A review paper identified research gaps and future directions for Agentic AI. The review paper identified critical research gaps and future directions in Agentic AI. Key priorities include designing reproducible evaluation protocols and developing multi-agent safety frameworks. The integration of online learning with long-term memory was also highlighted as a significant area for future research and development. SOURCE MATERIAL: (A Holistic Review of Agentic AI Frameworks, Applications, and Research Trajectories, 2026-06-15): Abstract Agentic Artificial Intelligence (Agentic AI), as a new generation of intelligent systems, extends beyond mere text or image generation by incorporating components such as multi step reasoning, persistent memory, multi agent interaction, and purposeful tool use. These features enable autonomy and dynamic decision making in open and complex environments. This paper provides an extensive review of the existing literature on Agentic AI. To this end, the conceptual and historical distinctions between Agentic Artificial Intelligence, classical agents, and general-purpose language models are first examined. Subsequently, the proposed architectural framework encompassing perception, role ad That is our update for this week. We will track more developments in our next episode. For now, thanks for listening to Research Agent.