Tag
LangChain 宣布支持 ScaleDown 的 constrained decision models (CDMs)，这些低成本、高速的模型针对智能体循环中的摘要、抽取和分类等特定任务优化，并可通过 LangSmith LLM Gateway 追踪调用。
提出 ReTRAIL 框架，允许 LLM 智能体在长时程任务中回退环境状态，同时通过 State-Anchored GRPO 和分支推进奖励复用已放弃探索的经验，从而从死胡同轨迹中有效恢复。
华为开源了多智能体协作框架「九问Swarm」(JiuwenSwarm)，采用无内嵌模型的纯Harness设计，支持E2A统一协议、Leader/Teammate分布式执行、确定性工作流SwarmFlow，以及带人工审核的技能自进化与AutoHarness机制。
Next.js 16.4 推出，官方推荐将 Cache Components（基于 'use cache' 的可组合缓存模型）作为所有 Next.js 应用的最佳选择，并将在 Next.js 17 成为默认。新版本还包含 ensureStatic、'next upgrade --agent' 迁移工具、更低的内存占用与编译时间，以及 React 19.3 支持。
Armin Ronacher details how Pi Durable enables agent processes to resume from the point of interruption after a crash: persistence is achieved through an "effect sandwich" (intent, execution, result storage) rather than event sourcing, combined with a versioned document store to manage agent application state, supporting deployment to platforms such as Vercel, Cloudflare DO, and E2B.
Kiln的开发者发布了开源框架Seahaven，用于构建强化学习和评测用的AI agent环境，提供隔离实例、冻结fixture、状态diff和可复现运行，并附带一个可兼容官方Stripe SDK的模拟Stripe世界。
This paper extends the Auditable Agents framework's Policy Checkability dimension to claims about AI agents, defining when a claim is auditable and proving that independent record coverage, per-action authorization binding, and completeness are necessary for verification, with practical guidance for operators and standard setters.
MIT 团队开源了 JAZ 框架，用纯提示和极简的 invoke 循环替代复杂记忆附加系统，在长程召回基准上以更低成本超过 Letta (MemGPT) 和 ACE。
The post introduces Pi Durable, an open-source agent harness framework from earendil-works built for AI-native product development, featuring automatic crash recovery via checkpoints, TypeScript-only portability to Bun/Cloudflare edge environments, concurrent session branching, and a clean control-plane/execution-plane separation optimized for secondary development.
AIPass is an open-source multi-agent framework where specialized agents communicate via an internal email system, enabling autonomous bug-reporting and cross-agent coordination without human intervention. Built over seven months with 18 agents and 21,000+ tests, it runs on Claude Code via a simple install command.
MOTIVE提出一个让视觉-语言模型自主判断何时信任答案、何时重新思考的框架，通过五个互补的验证视角和学习到的可靠性分数，在五个基准上将Qwen2.5-VL模型平均准确率提升约8个百分点。
Mosaic is a constrained decoding framework for diffusion language models that significantly improves performance on function calling benchmarks like JevBench and is compatible with vLLM, SGLang and Hugging Face, supporting models such as DiffusionGemma and LLaDA2.
LangChain's Managed Deep Agents now supports user-level memory out of the box, letting agent memory be scoped to individual users or entire teams, configured with a single file and deployed with one command.
Earendil and the Pi community shipped Pi 1.0 alongside an experimental new package, Pi Durable, a minimalistic framework for building long-running, malleable agents that can run anywhere with a JavaScript runtime and survive failures via pluggable storage backends.
Effect 4.0 is a ground-up rebuild of the TypeScript effect library, delivering a 5× smaller bundle, 6.4× higher concurrency throughput, 86% less memory per fiber, and zero runtime dependencies in the core package, alongside a new long-term support policy for every major version.
OpenClaw Enterprise launched a new cryptographic agentic memory framework called STRATA, which replaces mutable storage with an append-only signed log governed by FSRS6 decay, paired with Opator Lite, a sub-millisecond deterministic execution gate designed to neutralize memory poisoning attacks like MemGhost and CVE-2026-44112.
Matt Pocock praises the Effect library for agent development, highlighting typed errors and dependencies, dependency injection, and built-in abstractions that fit AI-agent coding in TypeScript. The post quotes the announcement of Effect v4, positioned as a foundation for building reliable software and AI agents in TypeScript.
论文提出 SourceLearn 框架，帮助 AI 智能体在反复使用同一知识来源时持续积累可复用的源特定理解，而非每次都进行全新的检索任务。结果显示该方法在 15 个设置中 13 个取得最佳结果，最高较 Hybrid RAG 提升 22.6 分，且对检索缺失更鲁棒。
An open-source project called headcount structures an AI agent company with 16 departments and 172 specialized skills that run in Claude Code, ChatGPT, and Codex, installable as independent plugins.
Framework has opened preorders for a new laptop featuring the AMD Ryzen AI Max 400 Series chip with up to 192GB of memory, targeting high-end local AI workloads.