Weekly Trending

Google AX is an open agentic orchestration runtime. It presents agent execution as a system of routes, execution boundaries, and oversight rather than an isolated chat interaction.

Anthropic's financial-services repository provides reference agents, skills, and data connectors for financial-services workflows, with outputs staged for human sign-off.

StarNet is a local-first pixel-art desktop harness that represents agent teams, capability-scoped permissions, workspaces, and runtime state as rooms and handoff lanes.

ResearchStudio-Idea is an evidence-grounded research-ideation skill suite that draws on machine-learning conference outcomes. It keeps source material close to the process of generating and assessing research ideas.

ResearchStudio-Reel extracts material from one research paper and turns it into editable poster, video, blog, and interactive-presentation artifacts, using a shared content spine and output checks.

EdgeBench studies how agents learn from extended interaction in real-world environments. It analyses roughly 38,000 hours across 134 long-horizon tasks and proposes scaling laws for environment learning after deployment.

LLM-as-a-Verifier treats verification as a scaling axis and uses fine-grained probabilistic scores for agentic tasks without extra training.

This work automates risk discovery, composes executable safety cases, and uses deterministic verification predicates for non-deterministic tool-using agents.

Qwen-UI-Agent describes a foundation GUI agent trained to use desktop interfaces through screenshots, pointing, typing, and a unified GUI/CLI action space. The report emphasizes a stateful execution harness and a data flywheel for real-world computer use.

A research system uses editable typed graphs and a coding-agent search loop to redesign an embodied agent’s perception, memory, planning, and action modules.

A study of why knowledge a language model has memorised can fail to route into the computations needed to solve a new task.

A multimodal museum assistant combines 2D and 3D artifact perception, external retrieval, source-level control, and reliability handling for uncertain evidence.

A survey of modern agentic systems that improve their behaviour by adapting components around the base model, including prompts, tools, memory, workflows, and evaluation loops.