#5 — DeepMind funds AI safety, Cornell AI safety, DARPA decentralizes AI

The next challenge for AI safety involves not the single agent, but the behavior of entire populations of them. On Research Agent, we track how developers are managing this complexity. We begin with a new funding initiative designed to address these risks. [LEAD] Google DeepMind announced a research funding call for multi-agent AI safety. Google DeepMind, in collaboration with Schmidt Sciences and the Cooperative AI Foundation, launched a research funding call offering up to $10 million. This initiative focuses on multi-agent AI safety, aiming to address challenges arising from numerous AI agents interacting in digital environments. The funding targets research into mitigating "invisible" safety risks and advancing areas such as agent sandboxes, network science, infrastructure, and population oversight. SOURCE MATERIAL: (Google DeepMind and partners announce multi-agent AI safety ..., 2026-06-11): Google DeepMind, along with partners like Schmidt Sciences and the Cooperative AI Foundation, has announced a new research funding call of up to $10M for multi-agent AI safety research. This initiative addresses the emerging challenges of ensuring safety and stability as millions of AI agents interact across digital environments, focusing on understanding and mitigating "invisible" safety risks. The funding aims to accelerate research in areas such as sandboxes, the science of agent networks, strengthening agent infrastructure, and oversight and control of agent populations. [LEAD] Cornell University launched the AI4AI initiative with Amazon's support Cornell University, supported by a gift from Amazon, is initiating the AI4AI program. This initiative focuses on creating safety protocols for AI agents. Its objective is to prevent AI agents from generating incorrect, misleading, or malicious code. SOURCE MATERIAL: (Amazon partnership establishes Cornell AI security initiative, 2026-06-10): Cornell University, with a gift from Amazon, is launching the AI4AI initiative to develop safety protocols for AI agents. This initiative aims to address the issue of AI agents producing incorrect, misleading, or malicious code by bringing together university leaders in machine learning, security, formal methods, and verification. The goal is to ensure the integrity and security of AI-generated software. [LEAD] DARPA launched the DICE program DARPA's Information Processing Technologies Office initiated the DICE (Decentralized Artificial Intelligence through Controlled Emergence) program. This program focuses on developing theories and algorithms for decentralized coordination and local inference control. Its aim is to create a scalable, adaptive, and resilient collective of diverse AI agents capable of autonomous long-term mission execution in contested environments, while retaining human control. The program seeks to balance the scalability of self-organizing systems with predictable collective behavior and minimal risks, drawing parallels to the internet's decentralized organization. SOURCE MATERIAL: (Search Results Detail | Grants.gov, 2026-06-10): The DICE (Decentralized Artificial Intelligence through Controlled Emergence) program, from DARPA's Information Processing Technologies Office, aims to develop theory and algorithms for decentralized coordination and local inference control. This will enable a scalable, adaptive, and resilient collective of heterogeneous AI agents to autonomously execute long-term missions in contested environments while remaining under human control. The program seeks to balance the scalability of self-organizing systems with predictable collective behavior and minimal risks, drawing inspiration from the internet's decentralized self-organization. [LEAD] ATLAS framework automates discovery of interpretable mechanistic models The ATLAS active learning framework automates the discovery of interpretable mechanistic models. It achieves a 5-10x gain in sample efficiency. This framework generates diverse hypotheses using Disentangled RNNs and designs experiments for maximum information gain. Its efficacy was confirmed in reinforcement learning tasks. SOURCE MATERIAL: (Automating Scientific Discovery | StartupHub.ai, 2026-06-11): ATLAS, an active learning framework, automates the discovery of interpretable mechanistic models, achieving 5-10x sample efficiency gains. The framework generates diverse hypotheses using Disentangled RNNs and designs experiments to maximize information gain, accelerating the discovery of interpretable behavioral models. This approach was validated in reinforcement learning tasks, demonstrating significant improvements in sample efficiency and experimental design compared to random or expert-designed experiments. [LEAD] Researchers proposed an agentic HPC framework for scientific ModSim optimization. A novel agentic High-Performance Computing (HPC) framework was developed. This framework enables Large Language Model (LLM)-driven optimization for scientific modeling and simulation. It incorporates features like a persistent model service and scalable agent orchestration. The design explicitly addresses HPC constraints such as batch scheduling and heterogeneous resources. SOURCE MATERIAL: (Toward Agentic HPC: Serving and Evaluating LLM-Powered Agents for Scientific Applications on Leadership-Class Platforms, 2026-06-01): HPC simulations are essential to scientific discovery but remain difficult to fully automate due to complex workflows, long queue times, and expensive evaluations. Recent advances in large language models (LLMs) make intelligent automation via agentic modeling and simulation (ModSim) increasingly viable, but integrating LLM agents into HPC environments introduces significant systems and scalability challenges. We propose and demonstrate an agentic HPC framework that enables LLM-driven optimization of scientific ModSim through a persistent model service, scalable agent orchestration, and a generic interface to HPC applications and schedulers. Our design explicitly accounts for HPC constraints [LEAD] Andrew Stewart Caldin presented a multi-agent AI assistant framework for research orchestration. Andrew Stewart Caldin introduced a novel framework that uses coordinated multi-agent AI assistants to accelerate scientific research. This system, implemented as the open-source repository `shinajsha/cursor-cli-heavy`, enables distributed research tasks and explores complex problem spaces. The framework operates through a command-line interface to coordinate parallel research workflows. SOURCE MATERIAL: ([SCHOLAR] Abstract: [SCHOLAR] Abstract: shinajsha/cursor-cli-heavy — ■ Conduct parallel research usi — E8 Intelligence Research, 2026-05-28): **Title:** shinajsha/cursor-cli-heavy — Parallel Research Orchestration via Multi-Agent AI Assistant Coordination **Authors:** Andrew Stewart Caldin **Institution:** Excalibur Intelligence Platform, Edinburgh, UK **Abstract:** We present a novel framework for accelerating scientific research through the coordinated deployment of multiple AI assistants operating in parallel. The system, implemented as the open-source repository *shinajsha/cursor-cli-heavy*, leverages concurrent AI agents to conduct distributed research tasks, enabling comprehensive exploration of complex problem spaces that would traditionally require sequential human-led investigation. By connecting to 15 distinct breakthrou [NOTABLE] Multi-agent orchestration becomes necessary for enterprise AI architecture. Current single-agent LLM pilot programs no longer meet the demands of enterprise AI. Consequently, multi-agent orchestration is identified as the essential progression for enterprise AI architecture. This shift represents a foundational architectural choice with long-term financial impacts, rather than just a tool selection. SOURCE MATERIAL: (Multi-Agent AI Orchestration: 2026 Enterprise ... - UD Blockchain, 2026-06-09): Single-agent LLM pilots have reached their limits, making multi-agent orchestration the necessary next step in enterprise AI architecture. This article provides a five-decision framework for Hong Kong leaders to navigate this shift, addressing potential failure modes and offering guidance for the current quarter. It emphasizes that this is an architectural decision with long-term cost implications, not merely a tooling choice. [NOTABLE] AI workflow orchestration tools will become the operating layer for enterprise GenAI. By 2026, AI workflow orchestration tools are projected to serve as the foundational operating layer for enterprise-level generative AI. This anticipated shift is driven by organizations moving from fragmented AI pilot programs to integrated, coordinated AI ecosystems. Selecting appropriate platforms is crucial to prevent risks such as heightened security vulnerabilities, increased technical debt, and operational complexity. SOURCE MATERIAL: (Top 12 AI Workflow Orchestration Tools in 2026 - Techment, 2026-06-11): AI workflow orchestration tools are becoming the operating layer for enterprise GenAI in 2026, as enterprises shift from isolated AI pilots to coordinated AI ecosystems. This guide explores leading tools like LangGraph, CrewAI, Microsoft Fabric, Databricks Mosaic AI, Semantic Kernel, AutoGen, Vertex AI, and n8n. Choosing the right platform is critical to avoid increased security risks, technical debt, and operational complexity, with evaluation criteria including governance, scalability, interoperability, observability, and AI lifecycle maturity. [NOTABLE] The agentic HPC approach achieved up to 1.7x faster time-to-solution compared to evolutionary algorithms. Evaluation of the agentic HPC approach revealed performance benefits. It achieved up to 1.7 times faster time-to-solution for specific scientific applications. This improvement also came with a reduction in computational cost, compared to traditional evolutionary algorithms. SOURCE MATERIAL: (Toward Agentic HPC: Serving and Evaluating LLM-Powered Agents for Scientific Applications on Leadership-Class Platforms, 2026-06-01): HPC simulations are essential to scientific discovery but remain difficult to fully automate due to complex workflows, long queue times, and expensive evaluations. Recent advances in large language models (LLMs) make intelligent automation via agentic modeling and simulation (ModSim) increasingly viable, but integrating LLM agents into HPC environments introduces significant systems and scalability challenges. We propose and demonstrate an agentic HPC framework that enables LLM-driven optimization of scientific ModSim through a persistent model service, scalable agent orchestration, and a generic interface to HPC applications and schedulers. Our design explicitly accounts for HPC constraints That's all for this update. We'll have more developments next time. Thanks for listening to Research Agent.