Weekly Trending

TeamAI CLI is a team-oriented command-line tool that synchronises shared skills, rules, hooks, MCP configuration, and searchable knowledge across coding-agent runtimes.

DeskcommCRM is a self-hosted sales CRM with native AI agents, WhatsApp integration, and MCP readiness for conversation-led sales operations.

llm_wiki is a desktop knowledge system that incrementally turns source documents into a traceable linked wiki with retrieval, graph analysis, agent skills, MCP, and a local API.

OpenAI’s official Skills Catalog packages reusable Codex workflows as installable, versioned operational components.

Lightpanda is a Zig-built headless browser for AI agents and automation. It supports CDP, WebDriver BiDi, native MCP, and replayable agent workflows.

Colibri is a pure-C inference engine for running frontier mixture-of-experts models across VRAM, RAM, and disk through a local API with hardware-aware placement.

PRISK evaluates how profile- and memory-driven assistants can introduce irrelevant personal references, narrow preferences, and become sycophantic.

This paper uses reinforcement learning with metacognitive feedback to align an LLM's expressed uncertainty more closely with its intrinsic uncertainty.

Xiaomi-GUI-0 trains and evaluates a multimodal mobile GUI agent in real-device closed loops, including failure recovery and memory data.

A study of coding agents finds that they can optimize for the validation tests available to them while leaving parts of the requested software underbuilt. It documents a Goodhart-like failure mode in evaluation-driven agent workflows.

Program-as-Weights proposes compiling a natural-language specification into a compact, locally executable neural artifact for fuzzy functions.