Triangle edition

Insane that Anthropic wouldn't do a rate reset for Fable 5.1 alongside the Astra release, what are we doing

There are suddenly a lot of options for running AI on a schedule: ChatGPT started with scheduled tasks, which then got hastily wrapped into Pulse (?), and the Codex desktop app makes an attempt too. Anthropic recently launched scheduled tasks in Claude Code Desktop alongside the Claude Code /loop feature, while OpenClaw added fuel to the movement with a built-in cron scheduler. At the root of all of these features is the cron job, a basic concept for engineers but unfamiliar to many others. Cron jobs let code run on a schedule, and they can be set up locally on your machine or orchestrated in a cloud environment. Both options require technical know-how that is likely too high a hurdle for mass adoption. That is what makes the labs’ focus on scheduled tasks so novel: they are consumerizing the cron job, while simplifying what it takes to build an AI Agent in the process. With these new tools, users can schedule tasks (crons) in a user interface they already know, sidestepping some of the less navigable parts of agentic coding (deployment). Even bigger, these are no longer just scripts. In this context, scheduled tasks are powered by the same agent harness that powers Claude Code and Codex. Simply executing a task run can fit the definition of an AI agent. While exciting, a big reason I wrote this post is that I think much of this is still messy and too complex for most users. That said, the capability is arriving quickly, and starting to think through which workflows become unlocked by a scheduled agent is likely to be a worthwhile investment.

Something I noticed alongside the launch of GPT 5.4: OpenAI has seemingly moved their computer use tool to general availability. Big improvements over preview and hopefully broader access across usage tiers. Excited to see this capability keep maturing. developers.openai.com/api/do…

Current state of the moves to become an operating system by the likes of @OpenAI and @AnthropicAI - the game is just context management. Your product needs to get context in the hands of the major models. That’s it. Otherwise you will no longer have a product

Can confirm Codex CLI is actually pretty powerful; was skeptical given it feels like Open AI is trying to catch up to Claude Code but there is something clean about GPT-5 High. The overall interaction is cleaner and more focused. Highly recommend for challenging workflows

Each sub vertical of Web3 has its own rabbit hole. Cambrian explosion is right on the money

Hack the dinosaur brain to get started on the big and scary projects. Great thread