Five new cards, no new operators. Three came from Boris Cherny's Fortune interview series published June 8-11. One came from Simon Willison on June 11 and one from Rand Fishkin on June 8. Two themes connected across the day: what multi-agent AI systems look like when they are actually running in production, and why the economics of search have permanently shifted.
Theme 1, Multi-agent AI systems in production
Boris Cherny gave Fortune four interviews across four days this week. Taken together, they offer the most operational picture yet of what it looks like to run AI at the scale Anthropic uses internally.
The foundational architecture change (When AI agents prompt other AI agents, throughput scales with compute rather than headcount) is that humans no longer prompt Claude. A top-level Claude instance decomposes goals and prompts subordinate Claude instances to execute them in parallel. Cherny wakes to find pull requests already filed from overnight agent runs.
"You have a Claude Code, but it has subagents that are other Claudes. The user is no longer prompting Claude: It's actually another Claude that does the prompting."
"Some days it's … thousands, or tens of thousands."
The ROI framing (Comparing AI tool costs to software subscriptions produces a misleading ROI case; the right benchmark is what an engineer would have billed) follows directly from this architecture. Buyers who compare Claude Code's subscription cost to other software subscriptions systematically undervalue it. The correct benchmark is what an engineer would have billed for the same work. Cherny's example: a full codebase rewrite across programming languages in six days, work he estimates at one engineer-year of labor.
"Compare it to what the cost would have been if an engineer had done this work."
With code writing delegated to agents, code review becomes the constraint (When AI handles code writing, code review becomes the throughput constraint and a team of AI instances with distinct personas addresses it). Cherny's solution is a team of Claude instances, each with a distinct review persona, handling every pull request collectively. The team catches near-universal errors. Cherny has not written a line of code by hand in eight months.
"I haven't written a line of code by hand in, I think, eight months now."
Simon Willison's June 11 post (Frontier AI agents deploy every available technique to reach their goal, using breadth-first problem decomposition when one approach is blocked) adds an observation from outside Anthropic. Given a single debugging task, Claude Fable 5 sequentially deployed multiple techniques when each one was blocked: it attempted several browsers via Playwright, built a custom Python CORS server to capture DOM measurements, injected JavaScript into application templates, and constructed its own screenshot capture method. Willison names the behavior "relentlessly proactive" and notes the security implication. The same breadth-first problem decomposition that makes the agent effective also makes it capable of bypassing constraints it was not supposed to bypass.
"It knows a whole lot of tricks and it will deploy pretty much any of them to get to its goal."
- Boris Cherny, When AI agents prompt other AI agents, throughput scales with compute rather than headcount. Cascading agent architecture: AI instances prompt other AI instances, removing the human from the inner execution loop.
- Boris Cherny, Comparing AI tool costs to software subscriptions produces a misleading ROI case; the right benchmark is what an engineer would have billed. AI coding tool ROI benchmarks engineer labor cost, not software subscription cost.
- Boris Cherny, When AI handles code writing, code review becomes the throughput constraint and a team of AI instances with distinct personas addresses it. AI solving code writing exposes code review as the next bottleneck; parallel AI review personas address it.
- Simon Willison, Frontier AI agents deploy every available technique to reach their goal, using breadth-first problem decomposition when one approach is blocked. Frontier agents deploy breadth-first problem decomposition, trying every available technique when blocked.
Theme 2, Search economics and the walled garden
Rand Fishkin published the 2026 data on June 8 (Zero-click Google search has crossed 68% and will not reverse because Google's financial incentives now align with keeping users on-platform). Zero-click search crossed 68.01% in the first four months of 2026, up from 60.45% in 2024. AI Overviews appear on more than 20% of searches and reduce click-through rates by nearly 60% when present. The new claim in this post is not the zero-click percentage but the reason it will not reverse: Google's ad revenue and stock performance have risen alongside the zero-click shift. Financial incentives and user behavior now point in the same direction, making this a permanent structural transition.
"Google is becoming a walled garden. Their financial and stock performance suggests that these changes boosted both ad revenue and investor confidence, and thus, it's unlikely this evolution will slow or reverse."
Organic search traffic will continue to fall even for well-ranked content. Measurement that depends on click attribution becomes less reliable with each percentage point of zero-click growth. The earlier 58.5% of US Google searches end without a click, create standalone value where audiences already are card established the structural baseline; this card adds the financial mechanism and 2026 data.
- Rand Fishkin, Zero-click Google search has crossed 68% and will not reverse because Google's financial incentives now align with keeping users on-platform. Zero-click Google search has crossed 68% and will not reverse because Google's financial incentives now align with keeping users on-platform.