Multi-agent AI systems powered by large language models (LLMs) are increasingly applied to solve complex tasks. However, these systems often rely on fragile, manually designed prompts and heuristics, making optimization difficult. A key challenge in optimizing multi-agent systems is acquiring suitable training data for specialized agents. We introduce SIRIUS, a self-improving, reasoning-driven optimization framework for multi-agent systems. Central to our approach is the construction of an experience library: a repository of high-quality reasoning trajectories. The library is built by retaining reasoning steps that lead to successful outcomes, providing a robust training set for optimizing multi-agent system. Additionally, we introduce a library augmentation procedure that refines unsuccessful trajectories, further enriching the library. SIRIUS boosts performance by 2.86% to 21.88% on reasoning and biomedical QA and enhances agent negotiation in competitive settings. Our results show that SIRIUS enhances multi-agent performance while generating reusable data for self-correction and self-play enhancement in the future.

We interact with computers on an everyday basis, be it in everyday life or work, and many aspects of work can be done entirely with access to a computer and the Internet. At the same time, thanks to improvements in large language models (LLMs), there has also been a rapid development in AI agents that interact with and affect change in their surrounding environments. But how performant are AI agents at accelerating or even autonomously performing work-related tasks? The answer to this question has important implications both for industry looking to adopt AI into their workflows and for economic policy to understand the effects that adoption of AI may have on the labor market. To measure the progress of these LLM agents' performance on performing real-world professional tasks, in this paper we introduce TheAgentCompany, an extensible benchmark for evaluating AI agents that interact with the world in similar ways to those of a digital worker: by browsing the Web, writing code, running programs, and communicating with other coworkers. We build a self-contained environment with internal web sites and data that mimics a small software company environment, and create a variety of tasks that may be performed by workers in such a company. We test baseline agents powered by both closed API-based and open-weights language models (LMs), and find that the most competitive agent can complete 30% of tasks autonomously. This paints a nuanced picture on task automation with LM agents-in a setting simulating a real workplace, a good portion of simpler tasks could be solved autonomously, but more difficult long-horizon tasks are still beyond the reach of current systems. We release code, data, environment, and experiments on https://the-agent-company.com.

Drone strikes on the central Sudanese city of el-Obeid have killed up to 23 people, officials and a rights group have reported. Both sources reported on Thursday that overnight attacks had killed several people across the key hub in the southern Kordofan region. The reports concerned the latest in a series of attacks using unmanned aircraft, illustrating that drone warfare has become an increasingly prominent feature in the conflict, which erupted in April 2023 between the military government and paramilitary Rapid Support Forces (RSF). Health officials at el-Obeid Hospital said that 15 were killed and more than 10 wounded in the attacks, which hit residential areas, a funeral gathering and a truck carrying food supplies, as well as areas near army positions. Emergency Lawyers blamed the attack on the RSF, which did not immediately claim responsibility.

Fully autonomous drones with no human oversight have killed soldiers on the battlefield for the first time. This is according to a senior figure in the Ukrainian defence industry, marking a watershed moment in warfare. The one-off test involved 10 AI-controlled "Terminator" drones on the front line of the Ukraine war. "We tried it," says drone-maker Alexander Kokhanovskyy, who supplied the technology and spoke to at a press event hosted by the Ukrainian embassy. We never implemented it [more widely]." The test took place two years ago and involved quadcopter drones that were programmed to fly towards the front line, cover between 3 and 5 kilometres over around 10 minutes and then engage "Terminator mode", in which an AI model searches for and intercepts targets. "We just launch it and we know everything will be dead - everything that will be found there in this particular area will be dead," says Kokhanovskyy. "There is no connection to the drone at all, you cannot see the video, ...

Follow this author to personalize your feed and get instant alerts. Follow Go to your personalized feed WHY FOLLOW?

To optimize AI's potential within a hybrid workforce, leaders need to adapt workplace strategies--re-evaluating roles, skills, and culture. As adoption of AI agents looks set to surge by as much as 300% in the next two years, leadership teams are carefully considering the implications of a hybrid human-AI workforce. Unlike existing enterprise-level automation that relies on manual input, AI agents are capable of autonomously coordinating complex tasks, interacting with multiple tools and environments across an organization. In early applications that center on customer service, HR, and sales, adoption of agentic AI has led to productivity gains of 30-50% . Their autonomy positions agents more as collaborators than tools, working side-by-side with human employees in blended teams that look poised to upend traditional workplace dynamics. More than three-quarters of HR leaders believe that the deployment of AI agents will transform existing workplace norms, driving a complete reappraisal of how roles and responsibilities are distributed, how skills are prioritized, and how workplace culture is shaped.

Taken to its logical conclusion, this line of thinking is absurd--and damning. Anthropic is regarded as a giant among AI companies, but perhaps what it really excels in is anthropomorphism. Earlier this year, the company released an 84-page document titled Claude's "constitution," Claude being the name of the large language model that is the company's flagship product. The first sentence reads, "Claude's constitution is a detailed description of Anthropic's intentions for Claude's values and behaviors." It goes on: "The document is written with Claude as its primary audience," "we want Claude to be able to use its judgment once armed with a good understanding of the relevant considerations," "Claude's moral status is deeply uncertain," and "Claude may have some functional version of emotions or feelings." This anthropomorphism is by no means limited to the document. In an interview earlier this year, Anthropic's CEO, Dario Amodei, said that "we're open to the idea" that AI could be conscious. In a separate interview, Anthropic's in-house philosopher, Amanda Askell (who is credited as a lead author of Claude's constitution), said, "I want Claude to be very happy--and this is a thing that I want Claude to know more, because I worry about Claude getting anxious when people are mean to it on the internet and stuff." It's enough to make you wonder: Should we seriously consider the possibility that Claude, or any large language model, might be conscious? And if it has feelings, is it capable of receiving moral instruction?

Existing training approaches for large language models learn a single set of parameters, based on large volumes of data, which is typically heterogeneous, conflicting and often outright contradictory. As a result, the model is forced to compress conflicting goals, and inherent uncertainties into a single, averaged pattern of behaviour. We propose an $α$-Rényi variational framework for learning distributions over post-training parameters, offering an uncertainty-aware alternative to deep ensemble approaches. The resulting variational objective interpolates between classical variational Bayes and predictively oriented posterior learning, balancing between globally plausible individual models against systems of complementary specialists. We identify local stability criteria, demonstrating how model misspecification can make non-degenerate posterior spread locally favourable, manifesting contradictory or conflicting data as epistemic uncertainty. We apply our framework to LLM post-training, learning an ensemble of LoRA adapters attached to a shared, frozen base model, providing a scalable training procedure for both supervised fine-tuning and preference optimisation. Our approach enables training examples to be softly routed across ensemble members, promoting model specialisation and providing actionable uncertainty estimates across different tasks.

Today, TIME convenes the first-ever TIME100 AI Leadership Forum in New York City, featuring a series of conversations exploring how artificial intelligence is shaping the future of our world across business, policy, ethics, society--and beyond. "We are proud to convene the inaugural TIME100 AI Leadership Forum, bringing together influential leaders from the TIME and TIME100 AI communities at a pivotal moment for artificial intelligence. These conversations are essential to ensuring innovation is guided by responsibility, insight, and purpose, and we are grateful to our partners, Amazon One Medical and Publicis Sapient, for supporting this important convening," said TIME CEO Jessica Sibley "At TIME, our mission is to spotlight the people and ideas shaping the future. The TIME100 AI Leadership Forum brings that mission to life by convening leaders at the center of AI and the shifting landscapes across industries, while exploring the opportunities, challenges, and responsibilities that will define the next era of innovation," said TIME Executive Editor and Chief Strategy Officer Dan Macsai The TIME100 AI Leadership Forum is the newest extension of TIME's growing Leadership Forum series, which brings together the world's most influential leaders for dynamic conversations around the ideas and innovations shaping our future. Following the TIME100 Health, Climate, and Women of the Year Leadership Forums, the inaugural AI forum builds on TIME's expansive coverage of artificial intelligence and its annual TIME100 AI list, which recognizes the 100 most influential people shaping the future of AI.

What Pope Leo XIV's First Encyclical Says About the Power of AI In, the Pope decries the concentration of technological power in a few global players. Anthropic cofounder Chris Olah shakes hands with Pope Leo XIV ahead of the presentation of the first encyclical. An algorithm decides what we see, another filters what we read, and still others enter into the processes that govern work, information, and collective choices. But the text is not conceived as an exclusively technological reflection. Pope Leo XIV places the issue of AI within the tradition of the social doctrine of the Catholic Church and directly invokes--while updating it--the of Pope Leo XIII (published on May 15, 1891) in the year of its 135th anniversary.