agent

作者发现 REA 结合算力舱与本地大模型（如 GLM 5.3 Flash、Qwen 3.8 27B）在一小时内成功逆向 WorkBuddy 读取微信聊天记录的机制，称赞其本地模型不会像在线模型那样拒绝，并开源了相关逆向方案。

The open-source project REA (Reverse Engineer Anything), which topped GitHub's daily trending chart, gives coding agents like Claude Code reverse engineering capabilities in the form of an MCP Server. With a single npx command, it can analyze and reconstruct source-less binaries, Electron apps, and APKs, returning results with a full evidence chain.

A video demonstration shows web browsing tasks handled entirely by the Clef-Flash 9B decision model running locally, extracting page text at runtime to drive decisions and completing tasks in 11-17 seconds, though performance drops on complex UIs.

推文热烈推荐开源的 DeepSeek Bot（基于 DeepSeek Harness 的插件），称其为“杀手级应用”，认为它将让国内一大批 Agent 软件（如 Workbuddy、Trae）失去竞争力，建议国内用户搭配 DeepSeek 桌面版使用。

Grok Bot's 'unlimited token 2.0' trick is shared, showing how to connect Grok Bot to a local model via Tailscale so it can run 24/7 on your own graphics card for effectively unlimited tokens.

CutBay 是一款 AI 视频代理，用户通过聊天描述需求即可自动完成规划、脚本、画面、配音、字幕和剪辑，支持从提示词、URL、文档、PDF、演示文稿等多种来源生成并迭代视频。

A Hugging Face blog walks through how a user built six custom small models in a couple of days by prompting the ML Intern agent, including a 0.8B CPU-friendly distillation of Qwen-Image 2.1's 9B prompt rewriter for just $16 in compute, and shares the prompting patterns (baseline checks, smoke tests, budget caps) that made it reliable.

A user vents frustration about the constant breaking changes in openclaw updates, ultimately abandoning the local agent tool in favor of hosted alternatives like Grok bot, Muse, and Dots despite the monthly cost.

Startup Hark, founded by serial founder Brett Adcock, has publicly released Hark Pro, an AI personal assistant built on a computer-use model that operates users' computers end-to-end while emphasizing privacy over ad-driven data monetization. The product features a custom interface with chat, proactive task feeds, and a visible browser-navigation window designed to build user trust.

Row-Bot 5.0 demo running qwen3.8:27b locally in Ollama on a single GPU, analyzing bills, a lease, and an insurance policy to spot a water-bill leak and an illegal rent increase, entirely offline with no cloud APIs.

Andrej Karpathy built an agent called autoresearch that runs AI research autonomously using just 3 files — you give Claude or Codex a small LLM training setup and the agent loops through try, commit, and train.

The author introduces a web crawler that minimizes LLM calls: by placing a small decision layer (Jev) in the middle of the Agent pipeline, it decides whether a link hop is worth making, only invoking Claude when full context and reasoning are truly required — significantly reducing the cost of using an LLM as a general-purpose if statement.

Clef Flash, a 9B model running at Q4 quantization on an RTX 5080, plays Google Snake in real time (within the game's 135ms turn limit) using only simple vision instructions, with no training or game-state hacking required.

UNIXAI Version 3.6 is a minimalist single-script bash harness that turns any local Ollama model into an autonomous conversational agent, featuring an endless reflection loop, message queue trimming, and built-in shell tool-calling.

LCU is an open-source tool that exposes Codex's original computer-use runtime outside the Codex app, letting agents in Pi, Claude Code, and other harnesses control the entire desktop with native app-approval panes and Touch ID-based permission management.

Vivix Labs introduces the Vivix A1 model and A1 Playground + Agent, enabling real-time AI characters with full-body movement, continuous listening, interruptibility mid-sentence or mid-motion, and persistent object interaction without scene resets.

An app developer announces the open beta of Agent Eros, an AI relationship coaching platform built on frontier models, and lays out the case for and against AI involvement in relationship advice.

Hcompany releases Holo4, a series of agentic AI models for generalist computer-use, available in 27B and 35B-A3B sizes, which perform well on benchmarks compared to frontier models.

The tweet comments on the alignment of AGY's pace with Japan's IT scene, highlighting a new planning mode feature in Google Antigravity 2.0 that allows agents to research and plan before implementation.

SR-Fraud is an outcome-supervised reflective LLM agent framework for non-stationary payment fraud detection, improving detection metrics over traditional methods on a production benchmark.