agent

Notion announces Ship OS, an agent-native feature that allows teams to run the entire product development cycle within Notion, from customer feedback to merged PRs, with agents handling triaging, routing, and summarizing.

The article analyzes how, after the release of AI Agents like Claude Code, software development job postings in the US have increased rather than decreased. It points out that the market is shifting from needing executors to needing senior developers who can steer Agents, and that Agents are redefining the role of programmers.

In an interview, Teacher Wang Yuquan overturns a consensus every twenty minutes, ranging from the Industrial Revolution to AI, discussing views such as the AI bubble might burst in 2029, Agent applications, etc.

This article describes four primitives for building reliable AI agents: evidence-gated memory, counted evals, one governance gate, and honest recursion.

Mispher is a device for Mac that enables dictation, rewriting, translation, and includes an AI agent for transcription.

Vantage is an open-source Windows app that uses an LLM to automate desktop tasks via natural language, enabling users to drive any desktop app with plain English commands.

Shubham Saboo announced the open-source Awesome LLM Apps repository (117,000+ stars), a collection of 100+ ready-to-run AI agent and RAG templates that developers can clone, customize, and deploy.

The tweet describes a pattern using Meta LOOP with Fable 5 as advisor, GPT-5.5 as orchestrator, and Gemini 3.5 Flash as worker, and converts it into an installable Agent skill.

Someone has open-sourced an Agent skill library containing 75 skills, compatible with Codex, Claude Code, Cursor, focused on web design, landing pages, animations, and WebGL. A standout feature is converting design screen recordings into detailed prompts, allowing Fable 5 to automatically generate HTML pages. It's free and open-source, you can freely fork and modify.

This paper introduces a proactive memory agent that runs alongside an action agent to prevent behavioral state decay in long-horizon tasks, achieving significant improvements on Terminal-Bench2.0 and τ^2-Bench. The authors also train Qwen3.5-27B using SFT and GRPO as an early step toward open-weight memory policies.

Shilong Liu proposes a taxonomy classifying self-evolving agents into artifact optimization, harness self-improvement, and model learning, providing a common language for emerging agent research.

NousResearch has launched Hermes Agent on the cloud, enabling users to deploy agents with two clicks and 60 seconds, with granular access controls and unified billing for teams.

The user released an open-source AI illustration skill called guizang-material-illustration, based on GPT-Image 2.0 and Codex Agent, which can generate 3D material explanation diagrams with Chinese labels, suitable for articles, weekly reports, PPTs, etc.

A developer built Marvis, an Android app that acts as a voice-controlled agent capable of operating other apps on the phone—tapping, typing, scrolling, and switching between apps—with interruptible task execution and safety confirmations. It is distributed outside Google Play due to restrictions on smartphone-use agents.

Proposes SkillReranker, an inference-time reranking framework that decomposes tasks and skills to build a directed acyclic execution graph and uses a cross-encoder for scoring, improving task performance and efficiency in agent skill selection.

SearchEyes uses a typed knowledge graph as the backbone of a simulated search world to unify training data, search environment, and reward signals. It proposes Perception-Knowledge Chains (PKC) for multi-hop path sampling and Hop-Anchored Policy Optimization (HaPO) for step-level credit assignment, achieving state-of-the-art performance on multimodal knowledge-intensive benchmarks.

NapMem is a framework that treats long-term user memory as a structured action space rather than passive retrieval, using a multi-granularity memory pyramid and reinforcement learning to train agents to navigate memory. Experiments show competitive performance on memory-intensive tasks.

ArtisanCAD is a skill-guided industrial CAD agent that uses expert-grounded knowledge distillation to convert ambiguous prompts into executable CAD procedures, improving performance on the Text2CAD benchmark and enabling editable B-Rep models for complex automotive components.

MiniMax M3 uses sparse attention to make long-horizon agents practical by keeping context cost predictable and low, enabling 500K-token contexts with minimal quality loss and significant speedups in production.

The paper introduces MedCalc-Pro, a new benchmark for evaluating LLMs in complex medical calculations involving single, multi, and nested calculator settings, along with an agent framework that improves performance through multi-tool selection and structured validation.