AI Agents

SKT is a verified data synthesis pipeline designed to teach language-model agents how to identify, apply, and coordinate reusable skills. Using public skills, it generates executable task trajectories for supervised fine-tuning and evaluation.

Mu is an open-source tool platform that exposes search, news, mail, weather, files, calendar and other services to AI agents through a single MCP endpoint. Its pitch is not just API aggregation, but a self-hostable service stack with real underlying capabilities.

Beacon reframes agentic visual reasoning around selective and effective tool use. Instead of rewarding more tool calls, it asks whether tools are necessary and whether they genuinely improve outcomes.

Microsoft Research’s Echoverse argues that synthetic environments for computer-use agents should not merely look like apps; they must behave like apps. Its co-evolution loop repairs environments while turning graded rollouts into model training signal.

Qwen-UI-Agent aims to unify mobile, desktop, browser, and DeepSearch environments under a single GUI agent framework. The report focuses less on simple screen clicking and more on reliable execution across real devices and long-horizon workflows.

Meta is framing personal AI agents as a foundation for its next wave of products. Mark Zuckerberg says billions of people could have agents that understand their goals and work on their behalf within five years.

Google is expanding Managed Agents in the Gemini API with Gemini 3.6 Flash as the default model, environment hooks, budget caps, scheduled triggers and free-tier access.

Perplexity is expanding its Personal Computer tool to Windows, aiming to turn local PCs into AI agents that can work across files, Office apps, Teams, and the web. The feature is rolling out to Max and Enterprise Max subscribers.

StateAct reframes computer-use agents around program state rather than screenshots. By letting the main agent inspect files, backends and DOM directly while delegating rare visual tasks to a GUI subagent, it improves benchmark results and reduces cost.

Multi-Head Latent Control proposes a lightweight way to guide LLM agents by reading hidden-state trajectories from a frozen model. Instead of relying only on prompts or external routers, it predicts whether to answer, clarify, call tools, abstain, or hand off to a stronger model.

Microsoft has introduced MAI-Cyber-1-Flash and Project Perception, two AI-driven security tools designed to find, assess, and help fix vulnerabilities. The company claims benchmark and cost advantages, yet the timing highlights unresolved concerns around autonomous security agents.

Microsoft CEO Satya Nadella argues that companies relying wholly on proprietary AI labs risk outsourcing not just infrastructure, but their own thinking. His proposed answer is to keep prompts, metadata, context, memory, and agent tooling under enterprise control.

Microsoft has introduced MAI-Cyber-1-Flash and Perception, a new agentic cybersecurity platform designed to help enterprises find, triage, and remediate vulnerabilities faster.

Cursor’s engineering team asked a swarm of agents to rebuild a SQLite-like engine in Rust using only the 835-page manual. The striking result was not just test performance, but the cost gap between different agent architectures.

AWS has released Loom in AWS Labs as an open-source reference platform for building and governing AI agents at enterprise scale. The project focuses on identity propagation, access control, deployment governance, agent registries, and human approval flows.

Andrew Ng has open-sourced OpenWorker, a desktop AI agent designed to move beyond chat and deliver completed work. It emphasizes openness, local-first storage, privacy, and model flexibility.

Industry speakers in Beijing showed AI moving into film production, commercial design, and 3D virtual sets. Academics in Hangzhou argued that truly self-improving agents are still at an early stage.

AREX introduces a recursively self-improving agent design for deep research tasks, using constraint-wise verification to guide targeted follow-up research.

OpenForgeRL addresses a practical bottleneck in agent development: powerful inference harnesses are hard to train end to end with open RL stacks. Its proxy-and-orchestrator design lets researchers collect trajectories from real harnesses while using standard reinforcement learning infrastructure.

NVIDIA Labs introduces NOOA, a model-agnostic framework that reframes agent development around familiar Python object-oriented programming. Instead of splitting behavior across prompts, schemas, callbacks, and workflow graphs, an agent becomes a Python object with methods, fields, docstrings, and type annotations.

Meta is expanding its AI chatbot beyond answers, images, and drafting into more assistant-like productivity features. The update adds calendar-based briefings, event planning help, ongoing tasks, and steerable research.