Four new cards, one new operator. Three cards come from the AI-native lane, where two independent sources converged on the same finding from different angles: domain expertise is the key multiplier for agentic AI, not coding skill or job title. One card comes from the game-building lane, where Zukowski published new data on Steam demo conversion that reframes why demos work at all.
Theme 1, Domain expertise is the variable that determines how much an AI agent actually does
Ethan Mollick's June 30 essay "The Twilight of the Chatbots" synthesizes recent evidence and names a transition that the numbers now support. Agentic AI inverts the AI-access argument: it amplifies domain experts rather than equalizing non-experts captures his core claim: chatbots helped non-experts fill knowledge gaps; agents amplify the output of domain experts who already know what good work looks like. The mechanism is specific. Domain experts specify goals precisely, catch errors before they cascade, and direct recovery. Expert sessions in Anthropic's data trigger 12 agent actions per prompt vs. 5 for novices, and generate 3,200 words of output vs. 600.
The Anthropic research Mollick cites has now been published in full by Zoe Hitzig, Maxim Massenkoff, Eva Lyubich, and colleagues, and is captured separately in Domain expertise, not coding proficiency, predicts AI coding agent success across every occupation. Across 400,000 real Claude Code sessions from 235,000 users, every major occupation lands within 7 percentage points of software engineers in verified success rates. Lawyers, researchers, and analysts succeed at agent-driven coding tasks at nearly the same rate as engineers. The research concludes: "domain expertise, and not coding proficiency, amplifies effective use of the tool."
These two cards arrive from independent angles. Mollick's is a synthesis essay making a frame claim about the shift in who AI benefits. Hitzig et al.'s is primary empirical research on 400,000 sessions. The convergence sharpens both.
- Ethan Mollick, Agentic AI inverts the AI-access argument: it amplifies domain experts rather than equalizing non-experts. The transition from chatbot-for-non-experts to agent-for-domain-experts, synthesized from June 30, 2026 data.
- Zoe Hitzig et al., Domain expertise, not coding proficiency, predicts AI coding agent success across every occupation. 400,000 sessions showing every occupation within 7 percentage points of software engineers; domain expertise as the primary success variable.
Theme 2, Long-horizon agent runs compress months of work to hours at a measurable token cost
The second Mollick card from the same June 30 essay (A 14-hour autonomous agent run completes 2 to 17 weeks of engineering work for $251 in token costs) lands a data point that makes the agent-era economics concrete: Opus 4.7 completed 2-17 weeks of human engineering work in a 14-hour autonomous run costing $251 in tokens. Fable ran 9 hours on projects that "would have taken a team well over a week." The mechanism is self-correction: the agent checks its own output, catches errors internally, and continues without human checkpoints, so the throughput ceiling is token cost per inference call, not engineer hours.
This sits alongside the commission-not-steer pattern from Mollick's earlier Fable essay (Frontier models capable of multi-hour autonomous runs require commissioning, not steering) but adds the specific economics that make it actionable. Teams now have a number to compare against: a sprint that costs $50,000 in developer time may cost $251 in tokens. The conditions and failure modes matter here. The economics break down when review and correction of agent output costs more than generating it.
- Ethan Mollick, A 14-hour autonomous agent run completes 2 to 17 weeks of engineering work for $251 in token costs. Opus 4.7 completes 2-17 weeks of engineering work in 14 hours for $251 in tokens; the cost ceiling is tokens, not labor.
Theme 3, Steam demos target a discovery cohort, not mass wishlisting
Chris Zukowski published survey data on June 30 showing median demo-to-wishlist conversion of 19.3%, with a 30th percentile of 15.9% and a 70th percentile of 23.2%. A 19 percent demo-to-wishlist conversion rate is the expected median; demos target hardcore discovery players, not mass wishlisting captures the mechanism behind that number. Demos attract "Sickos" - hardcore Steam players who play every demo to find games before the crowd. These players rarely wishlist but create word-of-mouth that reaches a broader audience. Optimizing for wishlist conversion misunderstands why demos work. The game is not for the mass audience watching the demo; it is for the small cohort of Sickos who will talk about it.
This pairs with Steam's Personal Calendar widget is now the primary pre-launch discovery path for indie games without large wishlist counts from June 25, which established that the Personal Calendar widget is now the primary pre-launch discovery channel for games that cannot reach the 100,000-wishlist threshold for Popular Upcoming. Together, the two cards describe the current Steam launch discovery model: demos seed Sicko word-of-mouth, Personal Calendar extends front-page time, and neither mechanism relies on raw wishlist count as the primary lever.
- Chris Zukowski, A 19 percent demo-to-wishlist conversion rate is the expected median; demos target hardcore discovery players, not mass wishlisting. Median 19.3% demo-to-wishlist conversion is healthy; demos target Sickos who drive word-of-mouth, not mass conversion.