The Next Two Levels of AI Adoption

Key claims

• The eight-level map ran out fast. Every published “the eight levels of AI adoption” in June 2026, from Level 1 (ask a chatbot) to Level 8 (a manager agent runs subagents). By November most engineers Taylor talks to run several long-running agents with an orchestrator, and the Codex and Claude desktop apps have brought orchestration to nontechnical teams. He presented the next levels at Thesis: 2027.
• Levels 1–8 increased delegation; 9–10 change how strategies are explored and tested. Both are experimental, token-hungry, and “working at a higher level doesn’t guarantee better results today.”
• Level 9, Simulation: make agent teams compete on strategy. You design an ecosystem, not a system; your job becomes “defining the goal, the scoring, and the rules of competition.” Good for growth experiments, pricing, content strategy, multi-solution engineering problems.
• Diversity needs synthetic people. Faceless subagents converge. Every’s evidence: Dan Lens, a prompt that replies like Dan Shipper, covers 26% of his actual points on 83 unseen threads with one guess and 86% with 100 guesses; KatePass, optimized by Jannik Jung from 594 of editor-in-chief Kate Lee’s editing passes (24,000 edits), makes about twice the good edits and a third the harmful ones of the rules-based version. Taylor also built a synthetic Slack of every colleague. Shipper’s point: “generation is cheap, but ranking is hard,” and a society of clones can argue its way to the better idea.
• Level 9’s limits: each team needs reliable orchestration (three competing orchestrators can mean three times the human attention), and the token cost only pays when a wrong strategy costs more than a real-world test.
• Level 10, Multiverse: rerun the world to filter out chance. The 2006 Columbia/Santa Fe Institute “music worlds” study (14,341 people, eight worlds, 48 songs) showed the same song could hit in one world and flop in another. Taylor’s Hacker News simulator (nearly 2,000 cloned commenters) predicted front-page headlines 60% of the time, and the misses were luck: one early upvote sent identical content down different paths. “A strategy that wins in 40 out of 50 is robust.”
• Level 10’s limit: if the personas or market model are off, averaging many runs “just gives you a very confident wrong answer.” It works today only on narrow systems like a single feed or focus group.
• Why the levels follow: each extends the time horizon and autonomy; token prices falling 90% make trading tokens for quality viable. The real question is when the economics tip for your own risk tolerance.

Notable quotes

“My answer, which I covered in my talk at Thesis: 2027 last week, is simulation: building a model of the world your agents work in, so you can test strategies before reality does.”

“You move from asking, ‘What’s the best plan?’ to asking, ‘What’s the best way to find out which plan is best?’”

“An orchestrator manages faceless subagents, which all converge on the same ideas.”

“The clones show what your colleagues might say, not what they will say.”

“Even if you model every individual well, you can still be wrong about the outcome.”

“A strategy that wins in one universe might just be lucky. A strategy that wins in 40 out of 50 is robust.”

“If Level 8 is a code factory for one project, Level 9 is multiple factories competing to find the best approach, and Level 10 is running those competitions many times to determine what works on average.”

The two prompts, verbatim

Level 9: “Using the persona files for each person on our marketing team, set up three synthetic versions of the team, each led by a clone of a different team member and told to plan our product launch their way. Have each team write a launch plan, then put all three in front of a panel of 50 synthetic customers built from our interview transcripts. Score the plans on how many customers say they’d sign up, have the two losing teams critique the winner and revise their plans, and run three rounds. Give me the final winning plan, what changed between rounds, and which objections from the team-member clones never got resolved.”

Level 10: “Run our launch simulation 50 times using the same customer personas, but vary the order in which they see the announcement and which early reactions they’re exposed to. For each of our three headline options, report how often it comes out on top, the range of outcomes, and flag any headline that only wins when it gets a lucky start.”

How it connects

• AI Agents: orchestration is Level 8; these levels sit on top of it.
• Every and Dan Shipper: the publication and the founder whose clone is the worked example.