Weekly Links: Anthropic's S1 lands, Exponential Time Hypothesis, and The Counting House

This week: Anthropic's S1 dropped to near-universal relief on the numbers; CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS lands (you can now let Claude erase all files without confirmation if you so wish...), and RIP Patrick Stewart.

On to the main stories:

• Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean? The TechCrunch article doesn't explain "What it means" well, but it does provide the background. This week, both OpenAI and Anthropic unveiled sample algorithms for a mathematical problem called 3-SAT that underpins the Exponential Time Hypothesis. 3-Sat is a complexity class of problems called "NP," which means solutions to the problem are "easy" to verify but very hard to find (the search space of possible options grows very fast as the number of variables increases). The Exponential Time Hypothesis was widely assumed to hold and stated that no algorithms could solve 3-Sat in polynomial time. However, both AI labs have found such (sub-exponential) algorithms. They are still not efficient algorithms, but if the proofs are ratified, this is still an important step. The proof also does not mean P ≠ NP (that all class NP problems have polynomial (fast) algorithms which would be a huge result.
• The Only SaaS Moats Left in 2026. I really like Rob Walling's end-to-end analysis of the current challenges for SaaS businesses. He calls out the challenges and remaining (and maybe new) moats. I would go a radical step further: this isn't just about exiting SaaS; it's about any subscription software business. Even if it is AI-powered. That's because now everyone can be AI-powered.
• The Counting House — Steve Yegge. Somehow Steve Yegge keeps writing at the cutting edge of software architecture. This article describes the evolved form of what many organizations are trying to build: workflows with embedded agents at scale. I like the agent welfare sections a lot (though I'm not sure they apply so cleanly to one-shot task agents). A key message: you need to think about AI workflows in a totally different way than workflows of the past.
• Anthropic says one AI agent beats many subagents now that models can handle complexity alone. This idea always runs and runs. Start the timer on when the messages flip back to "multiple agents" being better.
• Salesforce Expands Agent Fabric to Automatically Stop Runaway Costs and AI Agent Failures Before They Impact Customers. I don't write about our company, Safe Intelligence, a lot in this newsletter, but sometimes it's hard to resist. Proper evals and tracking for agents are critically important for enterprise deployments, and it makes sense that Salesforce would launch this. Only... organizations need this across all their agents, on any platforms, and you'd generally want your eval infrastructure to be independent of your hosting platform.

Wishing you a great weekend!