Steipete's monthly reminder that went mega-viral: 'You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents.' 13734 likes—the highest-engagement tweet of the entire week.

This is the evolution from prompt engineering to loop engineering. The human's job shifts from writing individual prompts to designing the automated workflows, checks, and feedback loops that keep agents productive over hours and days.

Ben Cherny (Anthropic) shares five tips for running Opus autonomously for hours/days: 1) Use auto mode for permissions so Claude doesn't ask for approval. 2) Use dynamic workflows to orchestrate hundreds/thousands of agents. 3) Use /goal or /loop to nudge Claude to keep going. 4) Use Claude Code in the cloud so you can close your laptop (easiest via desktop or mobile app). 5) Make sure Claude has a way to self-verify end-to-end: Chrome extension for web, iOS/Android sim MCP for mobile, start the full server for backend.

1865 likes—this is the practical companion to Steipete's philosophy.

Thibault Sottiaux (OpenAI) announces a 100-day challenge: every day they'll select one person who does impressive or incredibly useful work with Codex and give them 10X usage limits for a month. First one tomorrow. 6454 likes.

This is a clever go-to-market strategy disguised as a giveaway—it incentivizes users to push Codex to its limits and creates a stream of showcase stories.

Nikunj observes the vibe shift: from tokenmaxxing and token anxiety to token optimizing in just a few weeks. His hot take: companies should give copious amounts of token budget to employees to stay at the frontier and explore all the edges. Otherwise, it's way too easy to fall back to 'doing things how they have always been done.' 89 likes.

Rauchg reports Vercel AI Gateway recovers on average over 1T tokens a month—much like Stripe recovers revenue with smart retries on failed payments. Zero markup over the labs, adding redundancy, zero-data retention enforcement, observability, usage APIs, caps. 266 likes.

1 trillion tokens recovered per month means the reliability layer is already handling a massive volume of failed requests and making them succeed.

Levie addresses what the market got wrong about AI eating enterprise software. Building good software was very hard—AI has made it a bit easier, but it's still hard to build something with good taste, differentiated, high quality, secure. And that's only one component.

The plurality of costs in most enterprise software companies is actually on GTM, because at scale most categories are tough to break into and need heavy consultative selling and support. AI hasn't reduced that need—in many cases it requires it even more as landscapes get more complicated for buyers. If you make development cheaper and more abundant, the new problem of discoverability and market differentiation becomes the hardest part. 378 likes.

Levie on model stratification: use-cases have to stratify between model families in the next year or two. A split between frontier intelligence for high-end tasks and much cheaper models for high-volume workloads. Frontier will still be far bigger than today, but the low-end will get larger too.

The new hard problem: the layer that can efficiently route workloads to the right model becomes increasingly valuable. Agent orchestration that can cost-optimize while still performing the task successfully will be in a strong position. 141 likes.

RealMadhuGuru corrects a common misconception: training data is not low-skill grunt work. Labs need training data for high-economic-value tasks—complex, domain-specific knowledge built over years, spanning legacy tools that don't talk to each other.

That's why we have SWE agents and not knowledge work agents yet. Companies creating this training data (like Mercor) are doing extremely high-leverage, high-skill work. Critical to moving AI forward, and deeply underappreciated. 60 likes.

Garry Tan flags a growing bottleneck: educating people on how to use the AI tools has become a serious bottleneck. 543 likes.

Zara Yang explains why her Frontend Slides skill has grown organically: slides are inherently social. People see cool slides and always ask 'how did you make it'. People perceive those using HTML decks as more AI-native and AI-savvy. 88 likes.

This is an insight about AI product design: the most viral AI tools are those whose output is naturally visible and shareable.