AI Agents News and AI Updates

Latest AI Agents news, research, benchmarks, product releases, and industry adoption updates.

Latest Articles

• Anthropic says any lab can now let a language model agent run the whole protein design stack — Anthropic's Claude models autonomously ran a full protein design pipeline — installing and orchestrating existing open-source biology tools — achieving a 26.8% binding hit rate on novel minibinders, nearly double the industry benchmark of 10–15%, though independent replication is still pending.
• The Download: AI’s self-improvement problem, and what’s driving the heat — A new study finds AI agents still can't conduct open-ended research, casting doubt on near-term recursive self-improvement — and separately, OpenAI paused model work after its Astra model hit a "critical" safety risk threshold.
• When Agents Act on Web3: An Attack-Surface Survey of MCP, Skills, and Tool Calling — A survey of AI agent attack surfaces in Web3 finds that the share of MCP tools modifying external state rose from 27% to 65% — and current defenses stop fewer than 30% of attacks, while model-level safety refuses fewer than 3%. Four properties of blockchains (irreversibility, signing authority, continuous autonomy, sequence-level composition) turn normally recoverable agent failures into permanent losses.
• A Theoretical Framework for Parallel Lifelong MAPF Using Group Decentralized Planning — GD-RHCR extends the state-of-the-art RHCR multi-agent pathfinding algorithm with parallel group-decentralized planning, proving near-optimal guarantees while scaling throughput to higher agent counts at significantly lower per-plan computation cost.
• Collective Counterfactual Planning: Coordination, Consent, and Verification under Representational Constraints — A formal model called Collective Counterfactual Planning (CCP) proves that agent teams can solve problems no individual can, but goal completion is only valid when every requirement falls within the team's collective representational coverage — blind spots make verification illegitimate, not just hard.
• Wuying-Browser-Agent: Real-World Centric Fundamental Long-Horizon Browser Agents — Wuying-Browser-Agent-27B sets new open-source records on browser automation: 80.6% on WebVoyager, 66.7% on Online-Mind2Web, and 65.1% on BrowserBench (a new 350-task real-web benchmark averaging 37.9 steps per task) — through full-pipeline alignment spanning execution, training, and evaluation.
• LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents — LEGO-RL trains coding agents via policy-gradient RL in native harnesses (OpenHands, Claude Code, OpenCode), boosting Qwen3.5-35B-A3B on SWE-bench Verified by 4-9 points per harness — OpenHands 64.0%→70.4%, Claude Code 62.4%→68.2%, OpenCode 57.2%→66.6%.
• When AI Designs AI: Innovation or Imitation? — A large empirical study finds 96.8% of AI-designed AI methods recombine existing human algorithmic choices — nearly half exactly replicate a human design — and current LLM agents match human SOTA in only 10 of 72 tested configurations without reliable generalization.
• LLM-Derived Preference Judgments Are Not Self-Consistent — Across six LLMs tested on flight, apartment, and hotel preference scenarios, LLM-derived cardinal preference judgments show large, persistent inconsistencies — the same model gives contradictory willingness-to-pay estimates that cannot be reproduced by any single utility function.
• GxP-Agent: Process-DAG Topology for Reliable Clinical Trial Programming with LLM Agents — A DAG-structured multi-agent system for clinical trial programming achieves 100% structural match on CDISC-Bench — a task where all 11 single-shot attempts by five frontier models, including flat multi-agent approaches, score 0%.