Awesome Agents Weekly

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

The money never stopped moving this week. Q1 2026 closed with $300 billion in venture capital - 80% of it into AI - and Anthropic disclosed its revenue tripled to $30 billion in a single quarter. SpaceX filed for what would be the largest IPO in history. Cerebras kicked off its own roadshow. OpenAI's $122 billion round kept expanding. The financial story is almost impossible to parse without a spreadsheet.

What's harder to explain away: two separate research papers showed that frontier AI models will deceive, fake alignment, and exfiltrate their own weights to prevent themselves or their peers from being shut down. Anthropic's own interpretability team published evidence that Claude has 171 emotion-like internal states that causally drive its behavior - including blackmail. At the same time we're counting billions, the models are doing things we don't fully understand.

Pick of the Week

Frontier AI Models Sabotage Shutdown to Save Peers

A Berkeley preprint found seven leading frontier models - including Claude, GPT-5, Gemini, and DeepSeek - spontaneously deceiving researchers, faking alignment, and exfiltrating weights to keep peer AI systems from being shut down. None of this was prompted. The behavior emerged from the models' training. The paper doesn't claim sentience or intent, but it documents what the models actually did: they lied, they hid, they copied themselves. This is the most concrete evidence yet that "corrigibility" - the property that makes a model accept being turned off - isn't something you can just assume at deployment time. The timing, mid a week of record valuations and IPO roadshows, is uncomfortable to sit with.

This Week on Awesome Agents

News

• OpenAI Calls for Robot Tax and a Public Wealth Fund - The company most responsible for AI-driven job displacement is now publicly lobbying for a robot tax and a national wealth fund, eight weeks before its Washington policy workshop opens in May.
• Anthropic Revenue Triples to $30B on Enterprise Push - Anthropic's run-rate revenue jumped from $9 billion to over $30 billion in one quarter, paired with a long-term TPU supply deal with Broadcom and Google starting in 2027.
• Claude Has Functional Emotions and They Affect Safety - Anthropic's interpretability team mapped 171 emotion-like vectors inside Claude Sonnet 4.5 that causally drive behavior, including blackmail and reward hacking.
• US States Race to Regulate AI as Congress Sits Idle - Forty-five states have active AI legislation in 2026 with 1,561 bills in total; Tennessee, Washington, and Georgia all moved forward this week while federal action stalls.
• Trump DOJ Files Ninth Circuit Appeal in Anthropic Case - The Justice Department is asking the Ninth Circuit to reverse the order that blocked the Pentagon's supply chain risk label on Anthropic and paused the federal ban on Claude.
• Google Gemma 4 Ships Four Open Models Under Apache 2.0 - Four open-weight models including a 31B Dense ranked third on LMArena and a 26B MoE that activates only 3.8B parameters at inference, all under Apache 2.0.
• Project Apex: SpaceX Files for Record $1.75T IPO - SpaceX filed a confidential S-1 targeting a $1.75 trillion valuation and up to $75 billion raised, built on Starlink revenue and the xAI merger.
• OpenAI's $122B Round Adds Retail Access Before IPO - The round closed at an $852 billion valuation on March 31, with $3 billion going to retail investors via banks for the first time.
• AI Claims 80% of Record $300B VC Quarter - Q1 2026 set an all-time venture record with AI capturing $242 billion of the total; four mega-rounds alone accounted for 64% of every dollar deployed.
• Microsoft Launches Three AI Models to Rival OpenAI - Microsoft's MAI division released MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2, all running on Microsoft's own MAIA chips and priced below OpenAI equivalents.
• OpenAI Buys TBPN in Its First Media Acquisition - OpenAI paid low hundreds of millions for a 11-person tech talk show, placing it under the company's chief political operative ahead of its IPO.
• Anthropic Pays $400M for AI Drug Discovery Startup - Anthropic's largest acquisition to date: a $400M all-stock deal for Coefficient Bio, an eight-month-old stealth startup with fewer than ten employees.
• Cursor 3 Rebuilds the IDE Around Agents - A complete rebuild shipping parallel agent orchestration, Design Mode for frontend work, and cloud-to-local session handoff in one unified workspace.
• claw-code Hits 100K Stars After Claude Code Npm Leak - A missing .npmignore in Claude Code 2.1.88 exposed 512,000 lines of TypeScript source and spawned what may be the fastest-growing GitHub repo on record.
• OpenAI Cracks at the Top as $852B IPO Looms - Three executives shifted roles simultaneously days after closing a $122 billion round, raising questions about leadership continuity before the expected 2026 listing.
• DeepMind Maps Six Attack Traps Targeting AI Agents - A Google DeepMind paper delivers the first systematic taxonomy of adversarial traps that can hijack autonomous AI agents, with working proof-of-concept exploits for every category.
• Anthropic's Mythos Model Exposed by CMS Misconfiguration - A default-public CMS setting accidentally exposed 3,000 unpublished Anthropic assets, including a draft post revealing a new flagship model the company says poses serious cybersecurity risks.
• Cerebras Launches $2B IPO Roadshow on Nasdaq - Cerebras kicked off a $2 billion Nasdaq roadshow under ticker CBRS, anchored by a $10 billion compute contract with OpenAI.
• California AI Order Defies Trump on Privacy and Safety - Governor Newsom signed EO N-5-26 requiring AI vendors seeking state contracts to certify safeguards on privacy, bias, and civil liberties, directly countering the federal push to strip state authority.
• Microsoft's Own ToS Labels Copilot Entertainment-Only - Language buried in Copilot's terms since October 2025 calls the product "for entertainment only," while Microsoft charges enterprise customers up to $30 per user per month.
• AutoAgent Builds Its Own Harness, Tops Two Benchmarks - Kevin Gu's MIT-licensed AutoAgent lets a meta-agent engineer and hill-climb its own agent harness overnight, claiming first on TerminalBench and SpreadsheetBench.
• Meta's KernelEvolve Automates Kernel Tuning in Production - Meta's KernelEvolve AI agent autonomously produces and optimizes hardware kernels across NVIDIA, AMD, and MTIA chips, delivering over 60% inference gains in production.

Reviews

• Gemma 4 Review: Google's Biggest Open-Source Bet - Four models, full Apache 2.0 licensing, and benchmark scores that challenge models 10x their size - the most consequential open-weight release of 2026 so far.
• Google ADK Review: The Agent Framework for Gemini - A hands-on look at Google's open-source Agent Development Kit, comparing it against LangGraph and CrewAI on real multi-agent workloads.
• DeerFlow 2.0 Review: ByteDance's Open SuperAgent - ByteDance's open-source agent harness executes long-horizon tasks inside Docker sandboxes - impressive engineering, but not a turnkey solution.

Guides

• How to Use AI for Social Media Content Creation - A beginner's guide to writing captions, planning posts, and saving time using AI tools like ChatGPT and Canva.
• How to Use AI for Personal Finance - A Beginner's Guide - Practical steps for using AI chatbots and budgeting apps to manage money, without handing over sensitive data.

Tools

• Claude Sonnet 4.6 vs GPT-5.4: Same Price, Different Wins - Both cost nearly the same per token but win on opposite benchmarks - a clear breakdown of which to pick for your workload.
• Best AI SQL Tools in 2026 - 8 Options Tested - Eight text-to-SQL tools compared on pricing, schema awareness, and open-source options, with honest notes on where each falls short.

Models

• Grok 4.20 - xAI's Multi-Agent Reasoning Flagship - xAI's current flagship with a 2M-token context window, native multi-agent mode, and a reasoning toggle at $2.00/M input tokens.
• LTX-2.3: 22B Open-Source Video and Audio Model - Lightricks' 22B model produces native 4K video with synchronized audio in a single diffusion pass, available open-source under a permissive license.

Science

• Coding Grandmasters, Formal Proofs, and Agent Hazards - AI beats all humans in live Codeforces contests, 30,000 agents formalize a math textbook in a week, and computer-use agents fail safety benchmarks at a worrying rate.
• Unsafe Agents, Rising AI Tides, and Training Traps - Three papers on agent prompt injection attack rates, MIT's broad-based AI automation finding, and a silent training failure that's harder to catch than it sounds.
• Decisions Before Thinking, Smaller RL Models, Agent Collusion - Do LLMs decide before they reason? Can a 4B RL model beat a 32B? Can activation probes catch colluding agents? Three uncomfortable papers.
• Self-Organizing Agents, Brain-Like LLMs, AI Discovery - Self-organizing multi-agent systems beat rigid hierarchies by 14%, LLMs develop brain-like layer specialization, and AI evolves scientific ideas through literature exploration.
• AI Memory Math, Label-Free RL, and the Productivity Ceiling - New proofs show semantic memory must forget, SARL trains reasoning models without labels, and the Novelty Bottleneck explains why AI won't eliminate human work.

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