"The Next Two Levels of AI Adoption" - Mike Taylor (Every, Also True for Humans)
Why this is in the vault
Taylor extends Every's June eight-level ladder with Level 9 (competing agent teams) and Level 10 (rerunning the competition to separate skill from luck), which gives RDCO a vocabulary and a caution list for the "what's above orchestration" question.
The core argument
Levels 1-8 each increased delegation and autonomy; Taylor says the top is arriving faster than he expected, so the question becomes which strategy to bet on, not how much to hand off. His answer is simulation: model the world your agents work in and test strategies there before reality does. He flags both levels as experimental and token-hungry.
- Level 9, Simulation. Several orchestrated teams, each with its own strategy, compete toward one goal; losers get dropped. The human defines the goal, scoring, and rules. Faceless subagents converge on the same ideas, so he populates the society with synthetic people. Every's examples: a Slack-reply clone of Dan Shipper (first guess hits about a quarter of his real points, 100 guesses cover most of them), and an editing clone of Kate Lee tuned on hundreds of her real passes that beat the hand-written rules version. Limitation: three competing orchestrators need roughly three times the supervision, and tokens multiply. Human step: check the winning plan with the real people, since clones show what colleagues might say, not what they will say.
- Level 10, Multiverse. Run the same simulation many times with varied starting conditions to see what wins robustly rather than once. Evidence: the well-known 2006 "music worlds" study, where the same songs rose or flopped depending on early luck, and his own Hacker News persona simulator (about 2,000 clones, roughly 60 percent front-page prediction), where one early upvote sent identical content down different paths. A strategy winning 40 of 50 runs is robust. Limitation: a bad simulation averaged 50 times is a confidently wrong answer; it works today only on narrow systems like one feed or a focus group.
- Adoption stance. Falling token prices make trading tokens for quality plausible; the goal is knowing when the economics tip for you, and operators at Level 3 can ignore this for now.
Mapping against Ray Data Co
The most concrete hit is the investing stack: the investing-backtest-thesis skill and the paper-trade capital-cycle work already enforce multi-window, out-of-sample runs with confidence intervals, which is Level 10's "rerun to filter chance" applied to markets, and Taylor's "40 of 50 wins is robust" rule backs the existing ban on 2-decimal Sharpe from tiny trade counts. Second, the content engine: Level 9 maps onto Ship 30 / Sanity Check drafting, where several candidate angles could compete against a synthetic reader panel before the founder sees one, though the founder's "voice from real signal only" rule means any persona built from unpublished drafts is off limits; personas must come from published posts. Third, a gap and a contradiction check against the June note: that ladder concluded RDCO's real bottleneck is rung-5 reliability and review bandwidth, and Taylor's own limitation (more competing orchestrators means more human attention) says Levels 9-10 make that bottleneck worse, so this is a "watch, do not climb" item for the COO agent, not a roadmap change. For phData, it is useful framing for evals-minded clients: the "head of evals" author is selling simulation as the next eval surface.
⚠️ Sponsorship
House promo, no third-party sponsor. The issue ends with a pitch for Every Agent (Slack agent) plus the usual subscription bundle (Monologue, Compound Engineering). Taylor is Every's head of evals, and the Dan Lens / KatePass examples are Every's own internal tools with no outside validation, so treat the quoted performance figures as self-reported. The framework also nudges readers toward Every's consulting and tooling.
Related
- [[2026-06-02-every-eight-levels-ai-adoption]] - the Levels 1-8 ladder this piece extends; RDCO's rung-5 reliability conclusion still applies.
- [[2026-05-21-every-after-automation]] - Every's earlier "what comes after" framing of orchestration.
- [[2026-04-20-every-ai-autopilot-verification-decay]] - why more autonomous layers need calibrated verification.
- [[2026-07-20-every-ai-workflow-adoption]] - later Every adoption piece in the same series.
- [[feedback_voice_from_real_signal_only]] - constraint on building personas from unpublished drafts.