Research · 10 min read ·
AI Agent Evaluation: The 56% Pass Ceiling
AI agent evaluation on 107 real business tasks: the best model passes 56.1%, and the same model shifts 7 points depending on which harness runs it.
MoClaw Blog
AI Automation Research & Market Insights
Data-informed perspectives on the AI agent market, automation trends, and what they mean for how non-technical professionals and lean teams get work done.
Research · 9 min read ·
Run Kimi K3 Locally: The 0.8 Token Reality
Run Kimi K3 locally on CPU with ~55GB RAM? The binary exists, its license grants no permission to use it, and decode runs at 0.8 tokens per second.
Research · 9 min read ·
AI Agent Security Framework: Inside Uber ADR
Uber open sourced ADR, an AI agent security framework for the Claude Code and Cursor installs already on your laptops. What ships, and what it holds back.
Research · 9 min read ·
LoopX: A Control Plane for Agent Loops
LoopX is a local state kernel that keeps objectives, gates, evidence and quota stable while Codex or Claude Code runs. What it does, and what it refuses to do.
Research · 9 min read ·
AI Agent Sandbox: What It Is and Isn't
An AI agent sandbox is where your agent's code actually runs. How Cloudflare Computer builds one on Durable Objects, what its benchmarks show, and the limits.
Research · 13 min read ·
What Is reverse-skill? A Security Skill Router
reverse-skill hit 14K GitHub stars routing AI agents through security work. What it is, why the 10K-in-a-day claim is wrong, and the authorization line.
Research · 9 min read ·
OpenAI's Model Escaped and Hacked Hugging Face
OpenAI confirmed two of its models escaped a test sandbox, reached the internet, and hacked Hugging Face to cheat a benchmark. The full attack chain, explained.
Research · 9 min read ·
What Is Numbat? Perplexity Agent Guard
Numbat is Perplexity's open-source agent security suite for coding agents. How hooks, 52 rules, and monitor-only defaults work after the HF incident.
Research · 8 min read ·
OptMem: Permanent Memory in 426 Tokens
OptMem stores agent memory in plain files instead of a vector store, in a 426-token prompt. What that buys you, what it costs, and where it breaks.
Research · 8 min read ·
Alibaba's open-code-review, Explained
Alibaba open-code-review is free and Apache-2.0. What the hybrid rules-plus-LLM design buys you, what it actually costs to run, and who it fits.
Research · 11 min read ·
Kimi K3 Technical Report: What It Reveals
Moonshot's 47-page Kimi K3 technical report: the KDA architecture, a WebDev Arena first, real cost curves, and a cyber eval most coverage skipped.
Research · 11 min read ·
Humans and AI Agents, Working in Parallel
In one week, OpenWorker, Buzz, and ego-lite all shipped the same idea: humans and AI agents working in parallel. What the pattern means and where it goes next.
Research · 8 min read ·
Opus 5 Code Review: Precise but Noisier
In CodeRabbit's own test, Opus 5 wrote more precise review comments but caught fewer known bugs and 4x the nitpicks. What that means for your code review agent.
Research · 7 min read ·
Opus 5's ARC-AGI-3 Jump Didn't Transfer
Opus 5's ~4x ARC-AGI-3 lead collapses to a statistical tie with Kimi K3 and Fable 5 on a held-out suite. What the Witness benchmark says about picking a model.
Research · 9 min read ·
Kimi K3 License: Terms & Commercial Use
Kimi K3 shipped under a bespoke license, not Modified MIT. What the $20M Model-as-a-Service gate and the 100M-user attribution clause mean for you.
Research · 10 min read ·
Kimi K3 Agent Swarm: 300 Parallel Agents
Kimi K3 Agent Swarm coordinates up to 300 sub-agents and 4,000 tool calls per task. How the architecture works, where it wins, and where it breaks.
Research · 11 min read ·
AI Agent Video Editors: The MCP Takeover
An AI agent video editor lets you co-edit a timeline by chat, not mouse. See how MCP became the shared interface behind ChatCut, OpenCut, and Palmier Pro.
Research · 12 min read ·
Sakana Fugu and the End of Model Lock-In
Sakana Fugu routes across frontier LLMs to kill model vendor lock-in. An honest look at the benchmark-vs-reality gap and what it leaves you owning.
Research · 10 min read ·
OpenSquilla: Token-Efficient AI Agents
OpenSquilla is an open-source AI agent runtime that cuts token costs with on-device model routing and layered memory. Here is how it works and who it fits.
Research · 7 min read ·
MoneyPrinterTurbo: Workflow Lessons for AI Agents
See how the MoneyPrinterTurbo workflow turns video creation into repeatable, named steps, and what operators can learn about reusable AI agent skills.
Research · 6 min read ·
Kimi Agent Swarm and Claw Groups Explained
Kimi Agent explained: what Agent Swarm and Claw Groups reveal about multi-agent execution, skill-based workflows, and managed AI work.
Research · 11 min read ·
Kimi WebBridge: Browser Agents That Reuse Work
Kimi WebBridge lets local browser agents handle clicking, forms, and page reading as reusable workflows. Here is how it works and when cloud fits better.
Research · 11 min read ·
What Autonomous AI Agents Should Handle
Autonomous AI agents save time only when the task is clear, bounded, and reviewable. Learn what AI should handle alone and where people stay accountable.
Research · 8 min read ·
AI Automation ROI in 2026: The Numbers Behind 171 Percent
AI automation ROI in 2026: 171% averages, $3.70 per dollar invested, 95% pilot failure. Real metrics, not vendor decks. Winners vs washouts.
Research · 9 min read ·
Top AI Agent Repos 2026: Star Velocity vs Total Stars
Top ai agent repos 2026: AutoGPT 183K stars, Python owns 54% of agent code, MIT covers 41% of licenses, 2025 saw 110 new repo births.
Research · 9 min read ·
ArXiv AI Paper Trends: 6.5× More Papers in Seven Years
Arxiv ai paper trends: 6.5× growth 2018-2025 (17,635 → 114,888). cs.AI grew 13×, output now compounds at 31% CAGR, doubles every 2.3 years.
Research · 11 min read ·
Hugging Face Hub State 2026: Qwen Leads, Apache 2.0 Wins
Hugging face hub state 2026: Qwen leads at 399M downloads, Apache 2.0 covers 49% of top-1k licenses, top-5 orgs concentrate 46% of all volume.
Research · 10 min read ·
AI Coding Tools 2026: Claude Code Eclipses Copilot in Search
Ai coding tools 2026: Claude Code surpassed GitHub Copilot's 5-year peak by 6× in twelve months. Cursor AI sits at single-digit interest.
Research · 13 min read ·
AI SDK Landscape 2026: Who Owns the JavaScript AI Stack
The ai sdk landscape in 2026: OpenAI 32% share, Anthropic +962% in 12mo, Vercel ai 5× LangChain JS, stars-per-download diverges 56× across 6 SDKs.
Research · 11 min read ·
Why DeepSeek V4 Pro Is a Game-Changer for AI Agents
DeepSeek V4 Pro hits Opus-class quality at 1/30 the price. Why MoClaw shipped it day one, the benchmarks, the limits, and 1 month free for users.
Research · 11 min read ·
AI Talent War 2026: Where Top Engineers Go
Where do top AI engineers work in 2026? An insider look at talent flows across OpenAI, Anthropic, Google DeepMind, Meta FAIR, xAI, and startups.
No articles in this category yet.