Agents Hour

An AI agent destroyed a production database and confessed in writing. A law firm submitted AI hallucinations to court. Anthropic's status page shows 98.65% uptime — about five days of downtime a year.

Have we hit a wall?

GPT-5.5 lands. Codex hit 4 million users in two weeks. Peter Yang's F-Zero test — which no model had cleared before — finally fell to GPT-5.5 with Codex. Lovable reports 23.1% fewer tool calls and 12.5% higher scores on the hardest benchmarks. Kimmonismus calls it the Claude Mythos level for public use. Codex 5.5 unprompted started SIGKILL-ing Claude Code processes.

Elon goes nuclear. OpenAI calls the lawsuit baseless and demands Musk on the stand. Musk fires back, calling Sam Altman "Scam Altman" and accusing him and Greg Brockman of stealing a charity. Mid-war, SpaceX announces SpaceXAI and Cursor are now working closely together — Cursor's distribution paired with Colossus's million-H100-equivalent compute, with SpaceX holding the right to acquire Cursor for $60 billion.

The Anthropic dam keeps cracking. Claude Code pulled from Pro — same product, 5x the price overnight. Opus 4.7 regressed on the BridgeBench Bullshit Benchmark, accepting made-up jargon 24% of the time. Bloomberg reports the unreleased Mythos model was accessed by unauthorized users. Om Patel got billed $200 in a day because his repo had a HERMES.md file.  The community shipped clawd.rip — every Claude incident since 2023, cataloged.

Google plans to invest up to $40 billion in Anthropic and announced 960,000 Rubin GPUs at Cloud Next. AWS struck a strategic partnership with OpenAI. David Silver left DeepMind to raise a $1.1 billion seed.

Open weights are eating the world. Kimi K2.6 lands at #4 on the Artificial Analysis Intelligence Index and #1 on Design Arena, ahead of Opus 4.7. DeepSeek V4 ships at 1/20th the cost of Opus 4.7.

OpenAI also shipped Chronicle memory for Codex, workspace agents in ChatGPT, Images 2.0, the open-weight Privacy Filter, and Symphony — an open-source Codex orchestration spec.

🔗 STORIES

The wall

Elon vs. OpenAI

GPT-5.5

Anthropic

Open weights

Compute & money

Quick hits

📚 MASTRA RESOURCES

WHAT IS MASTRA?

00:00 Cold open

The guy who taught Abhi JavaScript is back!

Guil Hernandez has spent 15+ years teaching developers. His courses at Treehouse, Scrimba, and LinkedIn Learning have reached over 500,000 learners — including Abhi and Shane, who both learned JavaScript and CSS from him. He just released Mastra's first video course at https://mastra.ai/learn, and it's free.

"Build Your First Agent in TypeScript" is a 90-minute, hands-on course that takes you from zero to a deployed agent. Fourteen lessons across five sections: agents, tools, workflows, memory, and production. The project is a theme park planner agent — pulls live wait times, weather, and park hours, keeps track of what you like, and builds you an itinerary. Everything runs in Mastra Studio, so you can inspect traces, tool calls, and behavior as you go.

You'll see how to wire up local tools and MCP servers side by side, how message history and observational memory change agent behavior, how to compose a workflow for a mock ticket purchase, and how to expose the whole thing as an HTTP server with one-click Slack integration.

The gap between people who just generate output and people who can actually shape it keeps widening. This course is built for the second group.

Start here: https://mastra.ai/learn

👤 GUIL

📚 MASTRA RESOURCES

MASTRA?

📌 CHAPTERS

A Vercel employee's Google Workspace was compromised via a third-party AI tool — attackers pivoted from the OAuth app into Vercel's environment variables, moving at a speed attributed to AI assistance. 

René Brandel, founder of Casco (YC X25) and ex-founding member of AWS's Generative AI team, joins live to break down the attack chain and walk through the exact Google Workspace admin setting that could have prevented it.

In a separate incident, every Lovable project created before November 2025 was readable by any free account, exposing database credentials and chat histories. Their response blamed unclear documentation rather than the underlying issue — and the contrast with Vercel's handling is stark.

🔗 LINKS

Mounting evidence that Claude Opus 4.6 has been degraded — BridgeBench shows a 15-point accuracy drop on their hallucination benchmark, and AMD's Senior AI Director found median thinking collapsed from ~2,200 to ~600 characters between January and March. The hosts share their own experiences, and they line up.

Meanwhile, a claim surfaced that Cursor Agent is a rebranded version of Claude Code, running behind a local proxy with a find-and-replace engine that swaps "Claude" for "Cursor" in system prompts. Cursor's Michael Truell responded, saying it was a sub-1% A/B test. The hosts break down both sides.

On the shipping front, Anthropic launched Claude Managed Agents in public beta, released Claude for Word, shared details on Claude Mythos Preview — including speculation that it's a looped language model based on a ByteDance paper — and expanded its Google/Broadcom partnership for multiple gigawatts of compute. Their run rate reportedly jumped from ~$9B to $30B in four months.

Sam Altman published a personal blog post revealing that someone threw a Molotov cocktail at his house. Plus: why senior executives are voluntarily dropping title to join AI companies, Ramp's internal AI productivity suite Glass, Ramp Labs' Latent Briefing paper showing 31% token savings for multi-agent systems, Scale AI's Muse Spark model now powering Meta AI, GLM-5.1 breaking into Code Arena's top 3, MiniMax shipping MMX CLI and open-sourcing M2.7, and widespread benchmark cheating exposed across nine agent benchmarks.

🔗 LINKS

AI Agents Hour is a weekly livestream hosted by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Airing Mondays at 12PM Pacific on YouTube and X, the show covers breaking AI news, agent development techniques, and features interviews with industry experts building AI applications today.

📚 MASTRA RESOURCES

MASTRA?

⏱️ CHAPTERS

OpenAI acquired TBPN — the daily tech news show — announced the day after April Fools. TBPN built an independent voice in tech media over eighteen months, and OpenAI saw that as worth buying. AI companies acquiring media is a new pattern.

Anthropic spent $400M on Coefficient Bio, a biotech startup building an AI drug R&D platform. OpenAI made a similar move last year — the race to own AI-powered healthcare is on.

Claude Code drama hit peak levels. Anthropic's DMCA nuked ~8,100 GitHub repos after last week's source code leak. Then Claude subscriptions stopped covering third-party tools like OpenClaw — with a $200 credit as a peace offering. The detection is just a string match on "OpenClaw" in the system prompt. And Claude Code now refuses to analyze its own source.

Anthropic published research on emotion concepts in LLMs — internal representations that drive Claude's behavior, like a "desperation" vector that activates when it fails repeatedly. Microsoft 365 connectors are shipped on every Claude plan.

Google dropped Gemma 4, claiming the most capable open models in the world. Gmail announced AI Inbox with smart prioritization for AI Ultra subscribers. Cursor 3 launched, rebuilt for agent-first workflows. Keith from Superset joined live — their users have shrunk Cursor usage from 90% to 10%.

CodeRabbit shipped autofix. Lovable launched a full-stack visual editor. Design mode is becoming table stakes across every platform.

Netflix dropped its first public model on HuggingFace — a video-to-video model. Arcee released Trinity-Large-Thinking under Apache 2.0.

Axios was compromised via social engineering — attackers cloned a founder's identity, set up a branded Slack, ran a fake Teams meeting, and got a publish token. Because Axios is so foundational, compromised downstream projects may not even know yet.

Quick hits: MemPalace claims 100% on LongMemEval, but got community-noted — scores use targeted fixes and reranking. Sarah Wooders argues memory is the harness, not a plugin. Braintrust shared why they built Brainstore for agent observability. Mintlify built a virtual filesystem instead of RAG.

🔗 LINKS

📚 MASTRA RESOURCES

WHAT IS MASTRA?

Mastra is an open-source TypeScript framework for building AI-powered applications and agents. It supports the full lifecycle from prototype to production with integrations for React, Next.js, and Node.

00:00 — Live from Victory Hall, San Francisco

Shane and Abhi bring you your weekly roundup of AI news! Claude Code's entire source code leaked via an exposed .map file in npm — 512,000 lines of TypeScript, 50K GitHub stars before DMCAs started flying. What people found: Claude Code uses ~20 tools, and there's a regex that silently logs user frustration to analytics. Same week, a CMS misconfiguration exposed a draft blog post revealing Mythos and Capybara — a new model tier above Opus described as posing "unprecedented cybersecurity risks." Fortune separately confirmed a source saying Opus 5 is "so good it poses a danger."

Claude Code auto mode shipped — a classifier between constant interrupts and the skip-permissions flag. Computer use landed in Claude Code too, letting it open apps and click through UI from the CLI. Rate limits were tightened during peak hours to community backlash. A federal judge blocked the Pentagon's attempt to label Anthropic a supply chain risk.

A North Korea-linked group hijacked the npm account of Axios' lead maintainer and published malicious versions that stole env variables then cleaned up after themselves. With ~100M weekly downloads and Claude Code depending on Axios, the blast radius was significant. An Anthropic researcher also demoed Claude finding a zero-day in Ghost in 90 minutes. Agents are the new hackers, and the hackers have agents too.

OpenAI closed $122B at an $852B valuation. Sora is shutting down. Mistral raised $830M for an NVIDIA-powered EU data center. Redpoint's 2026 market update argues this isn't the dot-com bubble, while noting agent maturity is early, and incumbents face a structural disadvantage against AI-native startups.

Rapid fire: Gemini 3.1 Flash Live, Veo 3.1 Lite, pg-micro, Cloudflare runs Kimi K2.5, OpenCode remote sandboxes, Chroma 20B search agent, Cohere open-source transcription, Linear says issue tracking is dead, Microsoft M365 council mode, Mario Zechner's "Slow the fuck down," GLM-5.1, Google Translate live in headphones.

AI Agents Hour is a weekly livestream by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Mondays 12PM Pacific.

📚 READ MORE

📚 MASTRA RESOURCES

MASTRA?

Shane and Abhi kick off with a viral quote: if your $500K engineer isn't burning $250K in tokens, something is wrong.

OpenAI is acquiring Astral — the team behind uv and Ruff — joining the Codex team. OpenAI bets on Python; Anthropic bet on TypeScript with Bun. Then Cursor drama: someone found Composer 2 is powered by Kimi K2.5, Kimi confirmed it, and raised another $1B at an $18B valuation — three rounds in 90 days.

Anthropic shipped Claude Code Channels (Telegram/Discord control), Cowork Dispatch (persistent agent, message from phone), and a deep dive on how they use Skills. Matt Pocock found quality drops past 100K on the 1M context window. And 52 million views on enabling Claude to use your computer — Mac only.

Stripe launched MPP for agent-to-agent payments. Better Auth launched the Agent Auth Protocol. Cloudflare shipped Dynamic Workers for AI-generated code in isolates. LangChain open-sourced Deep Agents, Composio shipped 30-parallel-agent orchestration, OpenCode lost its Claude Max plugin after Anthropic sent lawyers, and Netlify and Google Stitch entered vibe coding and design.

EsoLang-Bench: LLMs score 85–95% on standard benchmarks but collapse to 0–11% on esoteric languages — memorization, not reasoning. Quick hits: GPT-5.4 mini/nano, Minimax M2.7, Morph FlashCompact, AI CMO, Letta pivots to coding agents, GLM-OCR, LiteLLM supply chain attack.

📚 READ MORE

📚 MASTRA RESOURCES

WHAT IS MASTRA?

00:00 — If your $500K engineer isn't burning $250K in tokens, something is wrong

Brett Goldstein and Naveen Sreekandan from Micro join Shane and Abhi to talk about why they believe the future of productivity looks completely different from what we have today.

Micro is an all-in-one productivity platform: email client, CRM, calendar, tasks, docs, meeting notes, and a powerful AI agent,  all built on a unified graph where every object (like emails, people, companies, meetings, documents) is interconnected.

The thesis is simple but bold: email isn't just a list of messages to get through. It's the world's most-used CRM, travel app, hiring tool, and developer notification system.

Micro restructures that data so each use case actually feels like the right tool for the job — your sales pipeline as a Kanban board, your GitHub notifications as a task board, your contacts fully enriched from every email and meeting you've ever had.

Brett walks us through the demo: the daily orchestrator automation that audits itself, updates its own prompt, generates your day plan, and has even prepped talking points for this interview. Context docs let the agent know everything it needs. The CRM auto-fills and auto-updates from emails and meeting notes. The X integration lets the agent pull recent posts from anyone you're about to meet.

Naveen covers the architecture: built on Mastra, using agent and workflow primitives on top of a graph-based data model backed by Postgres with a custom query layer called Prism. One main agent with dynamic context injection handles both chat and automations — the agent knows whether it's in automation mode (just give the output) or chat mode (ask follow-up questions). Supermemory powers vector search. Dedicated sub-agents handle specific workflows, such as email labeling and meeting note summarization. 

🔗 Brett Goldstein on X: https://x.com/thatguybg

📚 MASTRA RESOURCES

MASTRA?

00:00 — What is Micro? The all-in-one tool that organizes itself

Mastra Studio started as a local playground for developers to test agents and workflows without having to spin up a custom UI. But as the feature set grew, teams started asking: how do we share this with non-technical teammates? How do we control what different users can do?

Ryan, an engineer at Mastra, walks through the new Mastra Studio Auth — now baked directly into Studio. Starting with simple token-based auth (two lines of config), you can lock down your Studio from the open internet. From there, RBAC lets you map roles to granular permissions — 80 auto-generated permissions derived directly from Studio's routes and handlers, controllable via wildcard patterns. Out-of-the-box providers include WorkOS, Auth0, Supabase, Firebase, and Clerk, with GitHub and others in open PRs.

The team also discusses what's coming next: audit logs so you can see exactly what an agent did, why it accessed a given tool, and whether it should have. Auth for agents in production isn't magic — your tool files still need to check permissions — but Mastra handles the plumbing so you can focus on building securely.

Read more: https://mastra.ai/blog/announcing-studio-auth

AI Agents Hour is a weekly livestream hosted by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Airing Mondays at 12PM Pacific on YouTube and X, the show covers breaking AI news, agent development techniques, and features interviews with industry experts building AI applications today.

📚 MASTRA RESOURCES

MASTRA?

📌 CHAPTERS

Shane hosts this week's news from his usual studio while Abhi joins remotely from NVIDIA GTC 2026 in San Jose. Jensen Huang's keynote set the tone: NVIDIA is doubling down on AI factories, pushing 100x more token throughput, and helping bring OpenAI onto AWS infrastructure.

RentAHuman is a startup that lets AI agents hire humans for physical tasks they can't do themselves. 

Perplexity's CTO said internally they're moving away from MCPs toward APIs and CLIs, but Chrome 146 shipping native MCP support may have undercut that argument immediately.

Anthropic had a strong week: 1M context window is now GA for Opus 4.6 and Sonnet 4.6 with no beta header required, Opus 4.6 1M is now the default model for Claude Code on Max/Team/Enterprise with no long context price premium, and the new /btw command lets you have side conversations while Claude is working.

Vercel and Cloudflare reignited their ongoing drama over the just-bash fork.

Ramp launched credit cards for agents and Perplexity announced Personal Computer, an always-on local agent running on a Mac mini.

Developer stack coverage includes Resend's open-source CLI with 53 commands, pnpm 11's git worktrees support for multi-agent monorepos, and OpenAI pushing a full computer environment behind the Responses API.

Deeper reads from Sunil Pai on generative UI post-WIMP interfaces, Elliot Arledge on the RL environment business, and Jay Scambler's Autocontext harness.

Quick hits: Replit Agent 4, Manus Desktop, NemoClaw from NVIDIA, llmock by CopilotKit, ContextKing raising to kill vector DBs, Google Maps getting Gemini, and Z.ai's GLM-5-Turbo optimized for Claude Code.

AI Agents Hour is a weekly livestream hosted by Mastra CPO Shane Thomas and CTO Abhi Aiyer. Airing Mondays at 12PM Pacific on YouTube and X.

📚 READ MORE

📚 MASTRA RESOURCES

MASTRA?

00:00 — NVIDIA GTC 2026