framework

Search2o is an agent server combining an agent framework, dev environment, controlled runtime, semantic router, and encrypted cloud state management, supporting model-neutral LLM adapters and cross-agent conversation continuity.

NVIDIA introduces Cosmos Framework, an open, modular platform for training, post-training, optimizing, and deploying world foundation models for physical AI, covering RL pipelines like DPO, GRPO, PPO, and serving via Cosmos NIM microservices.

This paper proposes decomposing the agent harness into eight functional components and a portable, versioned specification framework that records which model ran in which harness, addressing score variability and attack-surface documentation gaps not covered by existing model/agent cards.

The authors announce Gear, an open-source, meta-learning-inspired framework described as "a scikit-learn for RSI," which separates search decisions from execution and evaluation so methods like GEPA, RRSI, and Harness² are easier to implement and compare. A working GEPA-style variant with shared failure diagnosis and staged evaluation showed pass-rate gains on AutomationBench and Terminal-Bench, plus a ~95% lower per-task API cost for GPT-5.6 Luna versus Astra using the same optimized harness.

An introduction to the open-source Python multi-agent orchestration framework Swarms (~7.3k stars), which supports sequential task handoffs, parallel execution, hierarchical team structures, and MCP and skills integration — designed to automate business processes into multi-AI collaborative systems.

Droid 1.0 is an agent harness that enables agents to run hundreds of tool calls in a single model request; building it to support multiple models ensures improvements generalize rather than overfitting to one model's quirks.

HarborRL has open-sourced a lightweight agent reinforcement learning framework that integrates model serving, trajectory generation, tool execution, verifier-based rewards, and policy updates into a reproducible pipeline. It supports flexible combinations of Qwen/GLM models × GRPO/DAPO algorithms × 200+ sandbox tasks × 40+ agent harnesses (such as Claude Code and SWE-agent).

PyTorch 2.15 发布，包含 3,117 次提交，重写的 NCCL 后端成为默认、新增 NIXL 单向传输、FSDP2 与流水线并行的更细粒度控制，并在 NVIDIA Blackwell（NVFP4）和 Apple Silicon（FP8）上带来更快的 LLM 解码，同时扩展了 AMD、Intel、ROCm 等平台支持。

Cursor 发布 Agent SDK，一个用于以代码定义、构建和部署 AI 智能体的框架，支持事件驱动的代码审查、合规工作流等场景，可通过一条命令部署到 Cursor 基础设施。

论文提出 LegoSR，一个统一的符号回归与微分方程发现框架，通过共享的方程中间表示整合七个求解器，并引入 LegoSR Agent 基于求解器知识和历史经验自适应选择求解器，在含噪 ODE/PDE 基准上接近七个专家调优求解器的并集性能，较通用 agent 最高提升 29 个百分点。

提出了 RSIHub，一个用于构建、比较和扩展 agent-evolution 方法的开源框架，通过可复用的算子与配方简化实验基础设施。论文在 Terminal Bench 2 和 τ³ Banking 上评测了 AHE、Hyperagents、A-Evolve 与 GEPA，最大 pass@1 提升达 13.5 个百分点，但方法排名随设置变化且优化收益不一定迁移到留出任务。

本文综述了数据、信息、链接数据及知识图谱的质量评估研究，提出包含24个质量维度和相应度量指标的知识图谱质量评估框架，以评估LLM时代知识图谱的可信度与适用性。

该论文提出了一个统一的“工作智能体”（work agents）框架，将基础模型与运行时框架结合以可靠执行复杂数字任务，引入持久性与开放性两个维度，并系统分析能力族、模型与框架优化及评估方法。

论文提出 SearchWeaver 框架，利用代理的执行轨迹和任务奖励自动联合优化语料表示、可执行工具与使用指令，从而改进 agentic search 的搜索接口设计。使用 GPT-5.4-nano 时，优化后的接口在 BrowseComp-Plus 上将答案准确率从 71.6% 提升至 78.4%，并在四个 BRIGHT 子集上将平均 nDCG@10 从 45.7 提升至 55.0。

PAPAV 提出了一个以能力为核心的统一框架（Perceive、Anticipate、Plan、Act、Verify），用于对比多模态具身智能体、计算机使用智能体与机器人系统，并基于 64 个基准揭示了 Anticipate 与 Verify 能力评估的显著缺口。

Harness² 是 Google Research 推出的即插即用递归 agent 框架，让 LLM agent 在无人工验证的情况下自动优化其执行 harness（提示、工具、记忆），在 WorkBuddy-Bench 等专业基准上最高提升 11.5 个点。

Google open-sourced Harness², a plug-and-play framework that turns any agent harness into a self-improving system by recursively proposing, probing, and composing harness edits without touching model weights. Tested with Gemini 3.5 Flash and GLM-5, it outperforms fixed harnesses in a single round and carries lessons across tasks.

一位CS学生分享了DenseCore——一个用Rust编写的轻量深度学习框架（约2MB），配有类PyTorch的Python API，能在Apple Silicon上逐比特精确复现PyTorch在nanoGPT上的训练结果，并揭示了PyTorch内部Flash Attention、Adam尾部循环融合等隐藏的数值细节。

Jim Fan's team open-sources V1.0 of ASPIRE, an agentic robotics framework with real robot skill discovery and self-improvement, noting frontier models like Astra, Opus 5.5, and Sonnet 5.5 have saturated their robot benchmarks.

Google Research 开源 Harness²，一个实现递归式智能体系统优化（recursive agent harnessing）的即插即用框架，可在模型权重固定的情况下通过 Propose–Probe–Compose 循环持续改进 agent harness，在多个专业基准上带来最高 11.8 个百分点的性能提升。