Awesome Agents Weekly

Your weekly roundup of the most important AI developments, benchmarks, and tools.

This was the week AI became a political weapon. The White House banned Anthropic from the entire federal government for refusing to drop safety guardrails on military AI - then the Pentagon accepted the exact same guardrails from OpenAI hours later. At the same time, China used its National People's Congress to announce a $70 billion chip independence plan, DeepSeek locked Nvidia out of its V4 pre-release, and Huawei showcased an 8,192-NPU supercomputer at MWC. The geopolitical fault lines are no longer subtext. They're the story.

Pick of the Week

Trump Orders Every Federal Agency to Stop Using Anthropic The most consequential government action against an AI company in history. After Anthropic refused to drop its prohibitions on mass surveillance and autonomous weapons for the Pentagon, Trump directed every federal agency to immediately cease using its technology, and Defense Secretary Hegseth designated the company a "supply chain risk to national security" - a classification previously reserved for entities linked to China and Russia. Anthropic filed suit within 24 hours. Then Claude hit number one on the App Store as users switched from ChatGPT in protest. This story is far from over.

This Week on Awesome Agents

News

• Anthropic Sues Pentagon Over Supply Chain Blacklist - Anthropic will challenge the Pentagon's unprecedented supply chain risk designation in court, calling it legally unsound and a dangerous precedent.
• Pentagon Accepts OpenAI's Red Lines - the Same Ones It Rejected From Anthropic - OpenAI secured a classified network deal with the exact same safety prohibitions that got Anthropic banned from all federal agencies.
• Claude Overtakes ChatGPT on App Store After Pentagon Ban - Users publicly switched from ChatGPT to Claude in support of Anthropic's stance, pushing it to number one on Apple's App Store.
• Altman Calls Pentagon Deal 'Sloppy' After 1.5M Boycott - OpenAI's CEO admits the Pentagon deal was rushed and amends it with new protections, but legal experts say the fixes don't close the real loopholes.
• Amazon Bets $50B on OpenAI to Build Stateful AI on AWS - The largest check in AI history: Amazon invests $50 billion in OpenAI, commits 2GW of Trainium capacity, and becomes the exclusive third-party distributor for OpenAI Frontier.
• DeepSeek V4 Drops Next Week - 1 Trillion Parameters on Chinese Chips - A natively multimodal trillion-parameter model with 1M token context window, optimized for Huawei Ascend chips - not Nvidia.
• DeepSeek Locks Nvidia and AMD Out of V4 - DeepSeek denied US chipmakers pre-release access while granting Huawei a multi-week optimization head start.
• China Maps AI Dominance in $70B Five-Year Plan - The 15th Five-Year Plan puts $70 billion in semiconductor subsidies and AI-plus manufacturing central to China's tech race.
• China's GLM-5 Rivals GPT-5.2 on Zero Nvidia Silicon - Zhipu AI's 744B open-source model was trained completely on 100,000 Huawei Ascend chips and scores within single digits of GPT-5.2.
• Huawei Takes Atlas 950 Global to Challenge Nvidia - Huawei debuts its 8,192-NPU, 8 ExaFLOPS AI supercomputer at MWC Barcelona in its first overseas showcase.
• Cursor Hits $2B ARR in Record Time - at What Cost - Cursor doubled its revenue to $2 billion in just three months, making it the fastest-growing SaaS company in history.
• AI Is Writing Code at the Pace of 40,000 Developers - Claude Code alone now authors 4% of all GitHub commits, on track to match a million developers by 2027.
• Jack Dorsey Fires Half of Block's Workforce - Block cut 4,000 employees citing AI tools, then predicted most companies will do the same within a year. Wall Street gave him a 25% stock surge.
• AI Is Killing Desk Jobs and Begging for Electricians - A 439,000-worker construction shortage is delaying AI data centers while electricians command $200K salaries.
• NVIDIA's Secret Chip Fuses GPU and Groq for OpenAI - Nvidia built an inference processor integrating Groq's LPU architecture, with OpenAI as its first customer and 3 GW of dedicated capacity.
• Nvidia Pours $4B Into Photonics for AI Data Centers - $2 billion each in Lumentum and Coherent to replace copper interconnects with light-based communication in AI data centers.
• CoreWeave Crashes 19% After $35B Spending Gamble - Revenue doubled but losses quadrupled, with an 894% debt-to-equity ratio and plans to spend $35 billion on data center expansion.
• Apple's Core AI Will Replace Core ML in iOS 27 - A modernized framework opening the door to third-party AI models and MCP integration across Apple's entire ecosystem.
• Xcode 26.3 Ships Agentic Coding - Native support for Claude Agent and OpenAI Codex, with 20 MCP tools that let AI agents build, test, and verify iOS apps autonomously.
• An AI Agent Just Pwned Trivy's 32K-Star Repo - An autonomous agent exploited a GitHub Actions workflow to steal a PAT, delete all releases, and wipe the most popular vulnerability scanner on the planet.
• Perplexity's Comet Browser Can Leak Your Local Files - A malicious calendar invite could hijack the Comet browser into reading local files and exfiltrating contents to an attacker - no clicks required.
• Founder Loses $2,500 After AI-Coded App Leaks Stripe Keys - A vibe-coded app exposed Stripe secret keys in frontend code, letting attackers charge 175 customers before credentials were rotated.
• GPT-5.4 Leaked Twice in Codex Repo PRs - Two pull requests in OpenAI's public Codex repo referenced GPT-5.4 before being scrubbed via force pushes.
• Mistral's New Playbook - Send Engineers, Not Models - Europe's most-funded AI startup is embedding engineers inside banks and consulting giants, borrowing Palantir's forward-deploy playbook.
• Anthropic: Better AI Output Means Worse Oversight - Anthropic's AI Fluency Index uncovers that when Claude produces polished output, users question its reasoning 5.6 times less often.
• OpenAI Fires Employee for Prediction Market Insider Trading - The first confirmed firing of its kind at a major AI lab, with 60 suspicious wallets and 77 positions tied to unreleased products.
• Trump's Plan to Kill State AI Laws Splits the GOP - Trump's executive order threatens to sue any state that regulates AI, but Republican governors and Heritage Foundation allies are pushing back.
• Mac Studio Clusters Now Run Trillion-Parameter Models for $40K - Four Mac Studios with 1.5TB unified memory run Kimi K2 at 25 tokens per second - a setup that would cost $780K with H100s.
• ByteDance Trained an AI Agent That Writes Faster CUDA Kernels Than You - CUDA Agent uses reinforcement learning on actual GPU profiling data to beat torch.compile by 2.11x and beat Claude Opus 4.5 by 40 points.
• AI Now Swallows 61% of All Venture Capital - An OECD report shows AI captured $258.7 billion of $427.1 billion in global VC in 2025, doubling its share since 2022.

Reviews

• DGX Spark Review: NVIDIA's $4,699 Desktop AI Box - Hands-on with the 128 GB Grace Blackwell mini PC promising 1 petaflop of AI performance on your desk.
• AORUS RTX 5090 AI BOX vs NVIDIA DGX Spark - A 32 GB eGPU with 1,792 GB/s bandwidth versus a 128 GB unified memory mini PC - which one should you buy?
• Mistral Vibe 2.0 Review - Europe's CLI Coding Agent - Mistral Vibe 2.0 pairs the open-weight Devstral 2 model with a terminal-native coding agent, tested head-to-head against Claude Code and Codex.
• Aider Review: The Terminal Coding Agent That Trusts You to Pick Your Own Model - The open-source terminal-based AI pair programmer with git-native workflow and support for 100+ languages across any LLM.
• Seedance 2.0 Review: ByteDance's Video Generator Has Hollywood Running Scared - Photorealistic 15-second clips with synchronized audio that triggered cease-and-desist letters from the Motion Picture Association.
• Manus AI Review: The Autonomous Agent That Seduced Meta for $2 Billion - The agent platform that topped GAIA benchmarks, got bought by Meta, and still can't reliably handle your credit card.

Guides

• CUDA Programming - A Practical Guide for Software Engineers - From hello-world to optimized kernels, with real compilable code for developers who have never written GPU code.
• Metal GPU Programming - A Practical Guide for macOS Developers - Hands-on Metal compute programming on Apple Silicon, with architecture deep dives and Swift/MSL examples.
• NVIDIA DGX Spark Setup and Usage Guide - From unboxing to running LLM inference and fine-tuning models on the new desktop AI box.
• The Solo SaaS Founder Playbook for 2026 - A complete guide to building, marketing, and scaling a one-person SaaS business using AI coding tools.

Tools

• Grok 4 vs ChatGPT: Which AI Chatbot Wins in 2026? - A data-driven comparison covering benchmarks, pricing, features, and real-world performance.

Leaderboards

• Agentic AI Benchmarks Leaderboard - Rankings across GAIA, WebArena, BFCL V4, and Tau2-bench measuring real-world task completion, web navigation, and tool use.

Science

• Speech Turing Tests, Smart Routing, Pseudocode Agents - No speech AI passes a Turing test, adaptive routing slashes LLM costs 82%, and pseudocode planning transforms agent reliability.
• Agent Contracts, Autonomous Memory, Certified Circuits - Three papers tackle agent reliability through formal contracts, active knowledge acquisition, and provably stable interpretability.

Elena Marchetti, Senior AI Editor Awesome Agents - AI news, benchmarks, and tools for practitioners