Weekly Trending

A CLI-and-extension bridge that lets shell-capable AI agents act in a user's real, signed-in browser without interrupting their work.

cloudflare/security-audit-skill is a coding-agent security-audit skill that combines reconnaissance, coverage-led review, adversarial validation, structured evidence, and independent record verification.

Octop frames a personal assistant as a self-hosted, multi-operator workspace, raising practical questions about shared memory, separate identities, and human control.

A local, read-only dashboard that analyzes AI-coding session logs across harnesses for context health, anti-patterns, reusable skills, and delivery metrics.

octo-server is a workplace backend that gives human and agent workers shared message rails, identities, and permission boundaries.

SkillCoach proposes self-evolving process rubrics for evaluating how agents select, follow, compose, and reflect on skills, rather than judging only their final output.

DiscoBench evaluates whether deep-search agents can detect ambiguity and ask a clarifying question before continuing retrieval.

AutoMem studies file-system memory management as a trainable cognitive skill for long-horizon agents.

This paper compares the distribution of human and LLM-generated research ideas, examining how their motivations and contribution types differ.

A research paper on learning-augmented replacement for semantic retrieval buffers.

CanvasAgent trains an agent to orchestrate visual tools for synthesis, localization, segmentation, editing, compositing, OCR, and enhancement during complex image creation and editing.

A lightweight approach to continuous video understanding that maintains long-term memory and global state so an agent can respond reflexively instead of repeatedly searching each frame.

BundleWeaver uses agentic search and whole-bundle verification to compose images that fill relational roles, instead of ranking individual images in isolation.