AI Agents

How would your agent stack change if a coverage might practice purely from its personal outcome-grounded rollouts—no rewards, no demos—but beat imitation studying throughout eight benchmarks? Meta Superintelligence Labs suggest ‘Early Experience‘, a reward-free coaching method that improves coverage studying in language brokers without massive human demonstration units and without reinforcement studying (RL) in the…

Meta AI has launched Agents Research Environments (ARE), a modular simulation stack for creating and operating agent duties, and Gaia2, a follow-up benchmark to GAIA that evaluates brokers in dynamic, write-enabled settings. ARE offers abstractions for apps, environments, occasions, notifications, and eventualities; Gaia2 runs on high of ARE and focuses on capabilities past search-and-execute. https://ai.meta.com/analysis/publications/are-scaling-up-agent-environments-and-evaluations/…

As AI brokers evolve past easy chatbots, new design patterns have emerged to make them extra succesful, adaptable, and clever. These agentic design patterns outline how brokers suppose, act, and collaborate to unravel complicated issues in real-world settings. Whether it’s reasoning via duties, writing and executing code, connecting to exterior instruments, and even reflecting on…

Sentient AI has launched ROMA (Recursive Open Meta-Agent), an open-source meta-agent framework for constructing high-performance multi-agent programs. ROMA constructions agentic workflows as a hierarchical, recursive process tree: dad or mum nodes break a posh objective into subtasks, cross them all the way down to little one nodes as context, and later combination their options as…

TL;DR: Computer-use brokers are VLM-driven UI brokers that act like customers on unmodified software program. Baselines on OSWorld began at 12.24% (human 72.36%); Claude Sonnet 4.5 now studies 61.4%. Gemini 2.5 Computer Use leads a number of net benchmarks (Online-Mind2Web 69.0%, WebVoyager 88.9%) however is not but OS-optimized. Next steps middle on OS-level robustness, sub-second…

TL;DR: AgentFlow is a trainable agent framework with 4 modules—Planner, Executor, Verifier, Generator—coordinated by an express reminiscence and toolset. The planner is optimized within the loop with a brand new on-policy methodology, Flow-GRPO, which broadcasts a trajectory-level consequence reward to each flip and applies token-level PPO-style updates with KL regularization and group-normalized benefits. On ten…

How do you audit frontier LLMs for misaligned conduct in real looking multi-turn, tool-use settings—at scale and past coarse mixture scores? Anthropic launched Petri (Parallel Exploration Tool for Risky Interactions), an open-source framework that automates alignment audits by orchestrating an auditor agent to probe a goal mannequin throughout multi-turn, tool-augmented interactions and a choose mannequin…

Which of your browser workflows would you delegate at the moment if an agent may plan and execute predefined UI actions? Google AI introduces Gemini 2.5 Computer Use, a specialised variant of Gemini 2.5 that plans and executes actual UI actions in a dwell browser through a constrained motion API. It’s accessible in public preview…