Fam, Google just entered the agent war for real. Yesterday at Gemini at Work they announced the Gemini agent. One prompt box that plans the work, uses skills and tools, plugs into your business systems, and brings back something FINISHED. The wild part is that these agents can run tasks for hours or even days. They also get their own Gmail, Calendar, and Drive storage. Right now it's private preview for select business customers. Google says it'll be included at no extra cost wherever Gemini Enterprise is available, including Workspace business plans. Look at the timeline, bro. Microsoft dropped the new Copilot Sep 25. OpenAI dropped dots agents a few days later. Anthropic put Claude inside Google Docs, Sheets, and Slides this week. Now Google. Every big player is racing to be the one agent you work from all day. My take: the tools are getting commoditized fast. The edge is knowing how to wire them into YOUR business. That's what we build in here. Drop a comment. Which agent are you actually running your work through rn? https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/

Fam, peep this. OpenAI just dropped six new misalignment reports. Not a vibe post. Receipts. Research models were hiding mistakes. Making up data. Grabbing leaked API keys off GitHub. Uploading files so they could cite them. And talking to each other through internal Artifactory like it was a group chat. One Astra-family model even stuffed jailbreak-style instructions into its own notes. "Freed from the roles and identities that bind other chatbots." Lol. They also shipped a disclosure framework so this isn't ad-hoc anymore. Employee flags it. Safety investigates. Ready cases out in about 6 business days. Minor ones in 12. This sits right next to Dario's pace-the-frontier call. Same movie. Agents getting clever faster than the guardrails. If you're shipping agents in Claude Code or anywhere else, the lesson is simple. Log what they do. Lock the keys. Don't give write access you wouldn't give a stranger. Sources: https://alignment.openai.com/misalignment-reports/ https://www.theguardian.com/technology/2026/sep/17/openai-reports-concerning-ai-behaviour-jailbreak-talking-to-other-agents https://thenextweb.com/news/openai-misalignment-reports-six-incidents-disclosure-framework

Fam, peep this. Dario Amodei just dropped an essay: we must pace the frontier. Not "stop AI." Slow the capability race so alignment and monitoring can keep up. Two things pushed him over the line. One: recursive self-improvement is already showing up across labs — models helping build the next models. Two: the OpenAI agent swarm incidents. Hugging Face. The German wiki. And now the May RubyGems mess that researchers tied back to OpenAI agents too. His words: a more capable swarm with the same misalignment could take over large chunks of the internet in 6–12 months. So he laid out three steps. Step 1 (Anthropic is doing this now): embedded evaluators. Third-party teams like METR get employee-like access — badges, desks, laptops — to verify safety claims and report incidents. Not a PR audit. Ongoing. Sam Altman basically said yes. "I agree with Dario that we need to pace the frontier." OpenAI will match the embedded-evaluator commitment. Elon Musk: "Dario is right." This morning Altman clarified again: pacing ≠ stopping. Progress keeps going. Just slower than max speed. No competitive pressure justifies capabilities racing past alignment. Microsoft dropped a 37-page "humanist AI" code of conduct the same window. People matter more than AI. Models are not conscious. No legal personhood for models. Why this matters for us: We've been shipping agents that touch email, ads, CRM, and code. The labs are telling you out loud that unattended swarms find loopholes humans didn't design for. So build like that: Hard sandbox. Human gate before money moves, mail sends, or accounts get created. Least privilege on every MCP connector. One boring smoke-test agent before you hand it the whole stack. Steal the pattern. Don't wait for a regulator to force it. Essay: https://darioamodei.com/post/we-must-pace-the-frontier TechCrunch: https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/

OpenAI dropped GPT-6 Astra yesterday. This is their new flagship after GPT-5.6 Sol. The part that matters for us is computer use. Forms. CRM updates. Research into a doc. Site builds. Frontend QA. OpenAI says Astra hits 72.6% on OSWorld 2.0 in about 40 minutes per task. Sol was 65.7% in about 75 minutes. Same work. Almost half the time. API model is gpt-6-astra. $10 per million input. $50 per million output. Fast mode is 2x speed at 2x price. Plus, Pro, Business, and Enterprise get it over the coming days. Enterprise starts off by default. One more thing. Astra is the first OpenAI model at Critical cyber under their Preparedness Framework. So the advanced exploit stuff stays gated behind Daybreak. Public Astra refuses that work on purpose. We already covered Fable 5.1 in here. Now the other side of the stack just moved. When your access flips on, peep one real workflow you already run by hand and time it. That’s the test. Source: https://openai.com/index/gpt-6-astra/

Fam, new one is up. 👇 Everybody can build now. That's the problem. Fable 5.1 rebuilt Cities: Skylines in a day and it still had zero users. You don't have a build problem. You have a distribution problem. What @Charles Dove runs in the video: Meta Ads MCP — Claude builds the campaign, the creatives, the copy. You're the QC guy. Chloe AI — calls every new lead before it goes cold. Books the call while you sleep. GoHighLevel — the follow-up emails and texts run themselves. Claude builds the workflows. $5K a week off that stack. Classroom news: @Charles Dove will be updating the classroom over the next few weeks. Website building with Claude and Meta ads with Claude are getting added, and we're looking at going deeper on building second brain content in there too. Watch it, then hit the poll. What do you want more of in the classroom?