subtl daily briefing

The details:

• ●Anthropic crossed the $1 trillion valuation threshold, surpassing OpenAI for the first time — a seismic shift in the AI power rankings.
• ●OpenAI restricted GPT-5.5-Cyber after the AISI (AI Safety Institute) flagged it as having record-level offensive cyber capabilities.
• ●Big Tech disclosed a week of massive AI capital expenditure figures, signaling the infrastructure arms race is still accelerating.
• ●A practical CLAUDE.md three-folder pattern (Root, Workstation, Project) has emerged as the scalable approach for building personal agents in Claude Cowork without terminal skills.

The details:

• ●Starbucks reversed its automation push and is hiring more baristas — a direct strategic bet on human-centered service as a differentiator.
• ●UChicago economist Alex Imas argues that when AI commoditizes execution, human presence and involvement become the scarce, high-value signal consumers pay a premium for.
• ●The move reflects a broader question emerging across industries: if AI can do it, does it still feel like enough to customers?

The details:

• ●Stanford and Google/MIT research found multi-agent systems amplify baseline errors by up to 17.2x compared to optimized single-agent setups when token budgets are controlled.
• ●Multi-agent architectures incur 2x–6x efficiency penalties on tool-heavy tasks — a cost that rarely shows up in demos but absolutely shows up in production.
• ●The recommended default: start with a single agent and fix reasoning gaps with pre-answer scaffolding prompts before ever escalating to multi-agent orchestration.
• ●Multi-agent systems are justified only when context windows collapse under noisy RAG data, tasks decompose into truly independent parallel sub-tasks, or strict regulatory verification (healthcare, finance) is required.