Five new cards, one new operator. No patterns: the convergences below involve two operators at most, which does not clear the threshold. Three themes emerged from this week's output: the cost economics of production AI, the safety and governance of agents running at scale, and the cognitive cost of handing too much to AI without friction.
Theme 1, Production AI economics are forcing a new evaluation metric
Tomasz Tunguz published two posts within five days that together describe a cost architecture shift happening at the enterprise level. In Enterprise AI buyers are shifting from benchmark rankings to intelligence per dollar as budget pressure reshapes model selection (June 3), he documents the shift from benchmark-first to intelligence-per-dollar as the primary model-selection criterion. The evidence is specific: Microsoft's MAI-Code-1-Flash matches Claude Haiku 4.5 on SWE-Bench at one-third the tokens; Uber hit its AI budget cap in four months; Salesforce allocated $300M for Anthropic tokens while freezing engineering hires. The buyer question is now "what is my intelligence per dollar?" not "which model tops the leaderboard?"
The companion card Frontier models can write procedural SKILL.md files that cheaper local models execute, decoupling reasoning cost from inference cost (May 29) describes one architectural response: use frontier models once to write procedural SKILL.md files, then run those files on cheap local models indefinitely. Tunguz frames it cleanly.
"The frontier model becomes a teacher. The library becomes the company's institutional knowledge. The student becomes whichever model happens to be cheapest this quarter."
Together these two cards describe a cost stack: measure outcomes per dollar, then design the execution tier to minimize the per-unit cost of those outcomes without degrading them.
- Tomasz Tunguz, Enterprise AI buyers are shifting from benchmark rankings to intelligence per dollar as budget pressure reshapes model selection. Budget pressure at scale is forcing a shift from performance benchmarks to intelligence-per-dollar as the primary model-selection criterion.
- Tomasz Tunguz, Frontier models can write procedural SKILL.md files that cheaper local models execute, decoupling reasoning cost from inference cost. Frontier models author procedural SKILL.md files once; cheaper local models execute them repeatedly, decoupling reasoning cost from inference cost.
Theme 2, Agent-scale security requires automated identity, not human gates
Two practitioners published independent arguments that human-gated security approval cannot survive the multiplication of AI agents. Simon Willison shipped datasette-agent-micropython on June 2 (Running MicroPython inside a WebAssembly sandbox gives AI agents safe code execution without host system access risk), demonstrating that MicroPython in a WebAssembly sandbox gives agents safe code execution, because WASM isolation prevents any escape to the host system. His test: "GPT-5.5 has so far failed to break out of the sandbox." The mechanism is containment at the execution layer.
Jonathan Jaffe (Lemonade CISO), speaking at a Theory Ventures event on May 28 (At agent scale, security governance must shift from humans approving humans to automated agent-identity and policy systems), argued for containment at the authorization layer.
"The future of security isn't humans gating humans — it's agents authorizing agents."
His framing is zero-trust applied to agents: each agent needs a unique auditable identity, and automated policy governs authorization at the point of action. Willison and Jaffe are solving adjacent problems in the same space. One addresses what the agent can execute; the other addresses what the agent can authorize. Neither is sufficient alone.
- Simon Willison, Running MicroPython inside a WebAssembly sandbox gives AI agents safe code execution without host system access risk. MicroPython inside a WASM sandbox converts code execution from a security risk into a safe tool call; GPT-5.5 could not break out in testing.
- Jonathan Jaffe, At agent scale, security governance must shift from humans approving humans to automated agent-identity and policy systems. Agent-scale authorization requires per-agent identity and automated policy at the point of action; human approval queues cannot scale to the interaction rate.
Theme 3, Frictionless AI extracts a cognitive cost
Ethan Mollick published "Choosing to Stay Human" on May 26 (Frictionless agentic AI triggers cognitive surrender, where people stop evaluating outputs critically even when the AI is wrong), arguing that agentic systems designed for frictionless delegation trigger what he and his Wharton colleagues call cognitive surrender: people stop evaluating AI outputs critically, even when the AI is wrong. The mechanism is the removal of corrective friction. Earlier AI tools required back-and-forth and frequent correction, which kept users engaged. Agentic systems remove that friction by design. Research on students and on knowledge workers both showed the same pattern: AI-assisted workers felt they were performing better and were not.
The card is a caution on the upstream end of the economics described in Themes 1 and 2. Cost-optimized execution is valuable. Frictionless delegation that removes human judgment from consequential decisions is not.
- Ethan Mollick, Frictionless agentic AI triggers cognitive surrender, where people stop evaluating outputs critically even when the AI is wrong. Agentic AI designed for frictionless delegation triggers cognitive surrender, where users accept AI outputs without evaluation, degrading both decision quality and learning.