Woke up to my Agents halted by a US Government order banning foreign AI usage. Best AI is now gate kept. We discussed this possibility, but not so soon. I'll say it again don't get locked into one tool, one model or stack you need be agent agnostic.

Last year, over 7 hours of tutorials I showed people how to build an app Levercast using sonnet 3.5. This morning Anthropic’s Fable built that app in 25min and $25. I’m gonna take the day off and dwell on this one. Where will we be this time next year?

Opus feels loose and creative, GPT feels tight and concise. This anthropomorphization means I lean on Opus for design and prototyping and GPT for the serious "don't mess this up" work. Is that just me? What models for what tasks?

Stop treating AI agents like employees; they’re temps. Every conversation is a new hire you have for one task (context window) In that light think of your agent temp in this way: - Onboarding (Docs, Context, Knowledge) - Tools (Skills, SOP, Search, MCP, API) - New Hire Overload (Be aware of context window) - Handover (What should they track so the next temp agent can pick it up)

Try using @cursor_ai or Codex with GPT 5.5 to generate images for your slides, app, or landing page based on your DESIGN.md style guide. You’ll get on-brand, color-matched images that make your designs stand out. It’s great to see this feature integrated into the IDE for agents

AI Promiscuous - Am I the only one planning for the future in this way, or is everyone happy being Agent Monogamous? The great thing about flexible development tools like @cursor_ai , Pi, @OpenCode : Changing models mid conversation when I'm not happy with the approach or opinion. Claude Code and Codex are great model suites but this is a strong reason why I don't want to get stuck in a single model ecosystem. Not to mention the fact, the new best model could come from anywhere and the cost model could change over night as we have seen. I don't want to lock my entire development team to one provider for something so important. The switching cost is low now, but when you factor in memories, workflows and skills AI lock-in becomes a larger problem than any legacy migration issue. Tell me I'm not the only one!?

Playing with Pi from @badlogicgames the last while. It feels like the Agent version of the classic VIM ide. It’s not for beginners but if you put some effort in you can get a very powerful experience without the lock in of Anthropic or OpenAI. We’ll be flexing Pimped Pi setups in a few weeks all over X #pimpmypi

You can build all your startup ideas with AI now, but here's why you probably shouldn't. Every success I've had, whether it's a company exit or growing my AI Builder YouTube channel to 40k subs, has come from going deep on a niche I'm passionate about (that also pays) for long enough to find the entrance wedge. AI helps us move faster, but going deep on a problem still matters more than ever. I started Reverbeo as a Twitter AI translation tool. It was only after months of working with our target customers that we realised what they actually wanted was a simple website translation solution. We were being too clever for our own good, but thankfully we didn't stop iterating. The lesson: pick a domain and iterate on ideas within it based on user feedback, rather than jumping to a completely different problem space every time something doesn't click. Same story with my YouTube channel, Rob Shocks. I started out making videos on Augmented Reality. They did okay, but it wasn't until almost a year in, when I started playing with AI agents which were adjacent to the subject, that the channel really took off. That's when I hit creator-market fit, teaching people how to build apps with AI, something I already had years of experience with. So in a nutshell: pick a problem space and domain you're genuinely passionate about (because staying motivated is harder than you think), start with your first idea, then iterate and pivot based on market response. Keep digging the same hole until you strike gold or rock. (More on how to spot the rocks in a later post.)

Composer 2.5 is actually really decent. I was a big fan of Composer 1 when it first came out, largely due to how fast it was, making it a joy to iterate with. Now, with its impressive intelligence thanks to reinforced learning and the massive compute resources from SpaceX, it's quickly becoming my default choice. Plus, it's 10 times cheaper! Love it! Who else is using it? What do you think?

Watch a team of humanoid robots running a full 8-hr shift at human performance levels. This is fully autonomous running Helix-02 x.com/i/broadcasts/1dJrPEVbZ…

Love, love, love

software engineers before vs after agents

Andrej Karpathy just shared an awesome talk about the evolution of software. He has a great way of coining new paradigms like Vibe Coding. In this talk: 1. Software 1.0: Hand-written rules. 2. Software 2.0: Neural network training. 3. Software 3.0: Using LLMs with prompts.

Stop chasing instant gratification with AI coding agents. Add features or bug fixes to your Linear or Notion backlog for later. It helps you prioritize and delegate work to agents instead of getting sidetracked by every little idea.

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

Cursor dropped version 3.1. My favourite feature is the ability to split tile your chat interfaces, allowing you to view multiple chats running simultaneously for each project. Check it out here: cursor.com/changelog

Introducing Claude Managed Agents: everything you need to build and deploy agents at scale. It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days. Now in public beta on the Claude Platform.

Claude Mythos - You Need To Prepare Now