AI Agents & Automation
-
fable51-worlds (PhiloLabs): Claude agents rebuilt SF’s Union Square — 453 buildings, 220 pedestrians, one prompt
Two days after Anthropic shipped Claude Fable 5.1, fable51-worlds (PhiloLabs) hit the HN front page with 248 points. It’s an open-source agent project: a swarm of fully autonomous Fable 5.1 agents researches a real place, models it, quality-checks itself, and ships a walkable Three.js world that runs in any browser. No game engine, no human… Continue reading
-
Perplexity Hybrid Compute sends your passwords to a local model, everything else to the cloud
Perplexity shipped Hybrid Compute for its Computer agent on Mac on September 1. Tasks run on frontier cloud models by default — Opus 5, GPT-5.6 Sol. An on-device classifier watches every step, and the moment it spots passwords, financial data, or health records, it asks if you want that step handled by a local model… Continue reading
-
AI agents wrote DoltLite in 2,000 PRs — a SQLite fork that passes 5.8 million tests
DoltHub needed a proving ground for Gas Town, Steve Yegge’s orchestrator that runs 20-30 Claude Code instances on one codebase. The result just hit Beta: DoltLite, a SQLite fork with the B-tree storage layer swapped for a Prolly Tree, built in 5 months by a team of AI agents across roughly 2,000 pull requests. What… Continue reading
-
Google Antigravity Boost (/boost) turns multi-agent reasoning into one slash command
Google Antigravity just gave paid users a nuclear option. Type /boost inside the IDE and it spins up a multi-agent deep reasoning pipeline for the problems a single agent keeps choking on — race conditions, hard algorithms, refactors that touch twenty files. How the pipeline works Three stages. A main orchestrator decomposes your task. Multiple… Continue reading
-
2,800 stars in a day: OpenMAIC (THU-MAIC) turns any document into a multi-agent classroom
Most AI education products are a chatbot with a syllabus bolted on. OpenMAIC, open-sourced by Tsinghua’s MAIC team, generates the entire classroom. Give it a topic or a document, and a multi-agent orchestration layer builds a full course: an AI teacher lecturing with voice and laser-pointer animations, AI classmates jumping into the discussion, agents drawing… Continue reading
-
OpenAI Site Tools (WebMCP in ChatGPT): Millions of Shopify Stores Are Now Agent-Native
AI agents have been using the web like tourists — screenshot, guess, click, hope. OpenAI decided that era is over. Site Tools, now live in the ChatGPT desktop app’s built-in browser, is the official implementation of WebMCP: websites expose structured tools, and ChatGPT or Codex calls them directly. No more simulating a human. How it… Continue reading
-
Construct Computer wins Product Hunt 1: your AI coworker gets a real cloud desktop
313 upvotes, first place on Product Hunt. Construct Computer is an AI agent platform built on one simple idea: stop trapping agents in chat windows. Each agent gets its own cloud desktop — browser, terminal, files, email, calendar — plus memory that persists across sessions. You log in from any device and watch it work.… Continue reading
-
Claude Tag update: Anthropic’s Slack agent now reads the whole conversation — and decides when to shut up
Anthropic pushed a major update to Claude Tag, its always-on AI teammate for Slack. The old version scored each message with a lightweight classifier, one at a time. The new one ingests the entire conversation — the back-and-forth, links, files, people — then picks one of four moves: reply inline, spin up a thread, route… Continue reading
-
Scientific Agent Skills (K-Dense-AI) hits 38,700 GitHub stars — 1,113 in a single day
Agent Skills started as a coding thing. K-Dense-AI just proved the format works for wet labs too. Scientific Agent Skills is an open-source library (MIT license) of 165 validated research skills plus connectors to 100+ scientific databases — cancer genomics, PK/PD modeling, molecular dynamics, RNA velocity, live pathogen-variant surveillance. Install it, and your generic AI… Continue reading
-
Microsoft ThinkingBox: 507 tasks, 20 runs each — the best AI agent is fully reliable on just 25%
Microsoft open-sourced ThinkingBox, a sandbox plus benchmark that tests whether AI agents can be trusted with real business workflows. The twist: it ignores what the agent says and only checks what actually changed in the backend database. One success isn’t reliability The benchmark ran 507 policy-conditioned tasks across five domains — retail, hospitality, auto insurance,… Continue reading
Top AI Product 