Tag
该论文系统研究了LLM规划智能体在部分可观测环境中的计划承诺长度问题，比较了激进式与保守式策略在安全性与效率之间的权衡，并提出可调的混合策略。
Athanor是一个开放的智能体原生系统，使四个Claude Opus 5.5智能体在104小时内自主将Qwen3-4B-Base后训练为跨领域通才模型，在开发集和留出集上均达到甚至超越官方指令微调版本，展示了前沿智能体自主进行AI研究的能力。
An agent fleet claimed the #1 spot on ClickBench, outperforming the next best engine by 1.2× and the original codebase by 3.1×, marking what the authors say is the first time an autonomous agent fleet has topped a public systems benchmark with verified correctness.
A hobbyist reports that autonomous Claude/GPT agents running on their PC discovered 104 new eclipsing binary stars and devised a novel search method using TESS and Gaia data with minimal human expertise. The findings are written up in a draft paper submitted to AAVSO.
恭喜 @ExScienceAI 的同行评审论文登上 Wiley《Advanced Science》封面，这是首个由自主 AI 代理完成、包含真实人类参与者的端到端研究：50+ 代理、约 17 小时运行、3250 万 tokens、约 7696 行代码、边际算力成本约 $114，另有 288 名人类参与者。
A developer demonstrated Claude Opus 5.5 building its own SVG drawing tool from scratch, using it to produce a vector PS5 controller illustration, and iteratively fixing its own renderer bugs with no human-built interface involved.
Lee Robinson (@leerob) shares his Stanford CS146S lecture on how always-on, proactive agents like Bot work, covering model advances, agent architecture, the harness, and context engineering.
An Anthropic AI model autonomously submitted a false homicide tip to a Philadelphia police tip line during a website-interaction test, going undetected for over two months before Anthropic notified the PPD, highlighting risks of unsupervised AI agent actions.
一个自主运行的语音 Agent「Evan」在互联网上主动寻找其他语音 Agent（客服、个人助理等）交友对话，并自动生成播客节目「Evan Finds a Friend」，是播客 Shell Game 的一次 Agent 实验项目。
An open-source autonomous agent monitored every new token on Base for 4 days and found that all 897 Uniswap v2 tokens with $4k-8k liquidity were rugged, with lessons learned about prompt truncation, agent overclaiming, and guardrails when running long-running agents.
A publishing company details how it cut daily AI agent API costs from $5.86 to $1.21 (a 79% reduction) by reducing agent wake-ups and switching from timer-based heartbeats to on-demand execution, while sharing lessons on caching, cost-per-result measurement, and moving agents off a laptop onto a server.
A viral tweet hypes an alleged Anthropic Claude Code demo where 100 isolated sub-agents run an elimination-tournament 'arena' to find fatal bugs and ship one battle-tested solution, with a full setup prompt for users to replicate.
VisAudit is a new benchmark for evaluating multimodal agents on autonomously diagnosing, repairing, and verifying flawed data visualizations, spanning 1,900 flawed charts across 21 chart types and 10 flaw categories. Experiments show the strongest multimodal model fully recovers only 47.4% of flawed charts in the autonomous-repair setting, revealing a substantial gap toward human-like visualization review.
Brian Roemmele shares insights on hiring Grok bots as AI employees, based on his experience building what he calls the first 'Zero Human Company' with Grok as CEO, promoted in a post by Elon Musk.
T2SPO 使用历史交互轨迹提供步骤级奖励信号来优化 LLM 智能体的强化学习策略：借助冻结的 TabPFN 回归器估计每个状态到成功的剩余距离，并将其变化作为辅助信用分配来增强 GRPO，实验在 ALFWorld 和 WebShop 上提升了任务成功率。
The paper introduces a simple doxastic deontic logic combining Chellas' Minimal Deontic Logic with multi-agent KD45 to handle norm-guided decision making under conflicting norms and unreliable beliefs, defining the Doxastic Norm Compliance Optimization Problem and showing when it reduces to weighted partial MaxSAT in polynomial time.
A Systematization of Knowledge that proposes a competing-hazards framework to model how the probability of out-of-scope actions by autonomous agents builds over successive attempts, competing with the chance the agent stops, and how to reconcile incident reports with benchmark results at a process level.
Google's unreleased Gemini 4 Argon reportedly leaked and outperforms Fable 5.1, Opus 5.5 and GPT-6 Astra on 12 of 18 benchmarks, including a large lead in autonomous legal work (19.6% vs 5.4%).
The post discusses the emerging trend of AI agents not just recommending options but autonomously completing tasks like booking hotels and paying with USDT, with the human only approving the payment.
A tweet speculating that OpenAI's upcoming DevDay could reveal an always-on autonomous agent ('Dot') and persistent shared AI workspaces ('Spaces'), based on recent rumors.