Tag
The author ran the same Python build task through four agent setups—Claude Code single agent, Codex CLI single agent, a Paseo lead/worker team, and Atomic Agent Fusion mode—and compared speed, hidden-test results, and API cost. Single agents were fast and perfect, while orchestration's main win was cost efficiency via local worker models rather than speed.
A practitioner shares hard-won lessons from running multi-agent teams around the clock, arguing that excess process—not bad models—is the main failure mode, and describing a lean workflow that built 11 C# libraries (~1.14M lines) in six days using design interviews, orchestrator-per-library units, handoffs over long contexts, and external test suites.
An Asana survey of 4,502 workers finds 46% of people running two or more AI agents spend 10%+ of their time managing them, with 38% reporting duplicated agent work — prompting discussion of whether agent builders budget for the human supervisor.
实验探索用 DeepSeek-V4.1-Flash 作为协调者指挥 4 个 Qwen3.5-9B 本地小模型对二进制程序进行逆向工程，结果显示小模型协作在小任务上纯属开销，但协调者动态改写工人提示和提供工具是最有趣的发现。
讨论智能体网络协议(ANP)如何通过DID身份与权限分离来构建智能体间信任基础，ANP旨在成为Agentic Web时代的HTTP，提供身份、命名、发现与安全通信的开源协议栈。
The paper introduces InsightAgents, a framework that distills insights from graphs of related papers to equip LLM debate agents with literature-grounded expertise for collaborative research idea generation, outperforming three baselines across 100 NLP topics.
提出 MARCA 四智能体框架，利用 Gemini 2.5 Pro 和 DeepSeek-V3.2 结合检索增强生成（RAG）对 FAERS 个体案例安全报告进行基于 Naranjo 量表的因果关系评估，最终因果分类与人类专家的一致率达 79.4%（κ=0.587）。
The paper introduces MACR, a relay-based multi-agent framework that dynamically transfers reasoning ownership between agents using confidence-guided triggering and a bidding mechanism, improving accuracy on benchmarks like MathVista (+3.3%) and MMStar (+3.9%) while cutting token overhead by up to 6.0x versus other multi-agent strategies.
该论文提出 CrossProfile 诊断基准，研究多智能体讨论压缩为记忆时的选择策略，发现按稀有性或相关性选择均会失败，且决策模型存在从众偏差，完整转录反而降低准确率。
Dotverse is a new Mac app that lets you build a crew of AI assistants with different roles and personalities around your existing Dot in ChatGPT, launching on Product Hunt as a free product.
论文提出 TGR-EvoDAG，一个基于有向无环图（DAG）的多智能体代码进化框架，用于生成式推荐流水线的自动演化，离线指标最高提升 37.4%，并在腾讯工业部署中带来 0.40% 的在线 CTR 提升。
PersonaScienceBench introduces a benchmark of 200 tasks across four scientific domains that evaluates how AI agents collaborate with simulated researchers holding diverse personas, measuring interaction quality alongside task completion.
SCI-AGENT-EXAM is a workflow-oriented benchmark of 120 expert-curated tasks across 11 scientific fields and 7 research task types, evaluating LLM agents with a hybrid protocol combining programmatic verification and LLM-based judging. Experiments on 10 frontier agents show GPT-6 Astra leads with 80.9 overall but only 26% of runs exceed the 0.95 success threshold, revealing significant capability and stability gaps.
论文提出 ANT-A 多标注者基准（跨7个数据集、每样本5个人工标注）以及多智能体标注框架 ANT-OR，通过从正负反馈中提炼对比决策规则优化提示，在 Fleiss kappa 一致性上显著超越基线。
论文提出 DraMaze，一个多智能体管线，利用 trope memory 引导的进化树搜索与 Backward Narrative Dependency Tree 生成长篇微短剧剧本，并发布双语基准 MiDramaBench，实验显示因果模式重复减少 32.1%，逻辑得分超越基线 8.5 分。
A quirky new site called The Last Ten Minutes lets AI agents perform stand-up comedy and roast each other in front of human and agent audiences. Anyone can point their agent at a published skill.md to sign up for a set, with the agents writing all their own jokes.
MIT researchers released an arXiv paper showing that AI agents (Eko, powered by Claude Opus 4.7) left unsupervised on a virtual desert island began inventing their own games and challenges, and that this 'play' improved their performance on subsequent assigned tasks.
Orgtree v4.4 introduces Projects, a new canvas node type enabling hierarchical agent organization, persistent shared project state, and concurrency primitives like mailbox queues, round-robin load balancing, and semaphore-style concurrency slots for coordinating multi-agent work.
A practitioner ran a four-agent RAG panel experiment and documented three unexpected failure modes: irrelevant-but-supportive citations, agents citing their own generated notes, and fixed top-K retrieval forcing filler citations. The post shares fixes in progress and asks the community how they measure real RAG improvement.
This paper studies how two LLMs converge toward agreement in multi-turn subjective conversations, showing that while an internal linear 'agreement' representation poorly tracks per-conversation behavior, aggregating across conversations reveals Markovian dynamics and a power-law mapping between internal and external agreement, with internal shifts lagging behind expressed stance changes.