In September 2025, we ended a ChatGPT training with a slide that read "become a conductor of knowledge & context," set over an image of a wizard waving a wand. We've taught that idea to thousands of people since. Ethan Mollick's new book, out this month, describes the next stage of AI work through a similar figure: the conjurer. We took it as a fun coincidence and read the chapter closely, and his description of conjurer work lines up with what we've been teaching.
The wand and the baton.
Mollick's Co-Existence: The Next Phase of AI (Penguin, 2026) describes three ways of working with AI. The third, conjurer work, is the one agents make possible: you summon AI workers and send them off with enough instruction to do the job. In his words: "In conjurer work, you define the task, assign it, and evaluate the output. You are not in the loop while the work is being done." (Co-Existence, "How to Work with AI," pp. 62-63)
Mollick is one of the best-known voices in AI. He teaches at Wharton, works with companies on how they adopt AI, and uses the tools heavily himself, and the book draws on all of it, along with experiments from the Generative AI Lab he co-directs. If you follow Mollick on X, you've seen parts of this argument develop over the past year.
Our sessions use a different image for the same concept: the conductor of knowledge and context. A conjurer sends workers off and a conductor directs an orchestra. In both pictures, the person no longer does each step by hand and still owns the result.
From executor to conductor.
The conductor idea starts with a change in what the job is. For most of our careers, knowledge workers have been executors: we wrote the report, built the model, and drafted the email. As AI takes on more of the execution, the person's job becomes moving knowledge and context to the right place so the AI can do that work well. We described this shift from executor to orchestrator in Focus vs. Multitasking. As people get better at it, they start running several streams of work at once, directing each one without playing any of the parts themselves. We first used the phrase "conductors of knowledge and context" publicly in an October 2025 post building on Dan Shipper's idea of the allocation economy. In January 2026, The Great Compaction described work consolidating around the people who can orchestrate several AI workstreams at once.
Mollick's conjurer follows the same progression. He contrasts it with the two modes he introduced in 2023, centaurs and cyborgs, where you do the task alongside the AI. Those modes don't scale, he argues, because you can only be in so many loops at once (pp. 62-63). The conjurer treats AI as an independent worker rather than a tool you operate, which he calls the most powerful mode and the most dangerous. He recommends it for routine or long tasks you don't need to watch (pp. 67-68).
Our conductor article describes where this leads: models calling other models, agents starting other agents, work running on a schedule, and a person deciding which agent handles which step and when a result is finished.
Both grew out of the end of prompt engineering.
We arrived at the conductor by way of context engineering. In Prompt Engineering Was Step One. Managing Context Is the Job Now., we argued that results depend more on the context a model works from than on how a prompt is worded. The conductor article described conducting as context engineering in practice: bringing the right knowledge and context to the work at the right time. We put it more simply on X that same week: prompt engineering quietly stopped being the skill, and what replaced it is conducting, deciding what context to put in front of the model.
Mollick reaches the same place, with research to back it up. His team tested politeness, insults, threats, and chain of thought instructions across thousands of questions and found almost no effect on the newest models, because larger models are less sensitive to wording and the techniques that helped have been built into training (pp. 53-54). What remains is delegation: "We have moved from a world of prompt engineering to a world of delegation." (p. 67) He says delegation depends on clear instructions, well-defined tasks, and careful evaluation, and he compares a good handoff to a product requirements document, a military five paragraph order, or a creative brief.
Each of those is a context document. A creative brief gives someone the goal, the audience, what done looks like, and what to avoid, which is the knowledge and context a conductor puts in place before the work starts. Mollick also suggests that when an AI finally succeeds at something it first failed at, you have it write down how in a text file for future AIs. That is what skills are: saved context, so the next run starts from what worked.
Both come at a cost.
Mollick says plainly that directing AI workers can be exhausting, because the old limits on how much one person could take on no longer hold (p. 67). He cites a study of workers at a technology company who became much more productive and also more burned out, because they no longer knew where work ended. His advice is to set careful boundaries around your conjuring. He adds that less experienced workers are hurt the most, like the sorcerer's apprentice (pp. 67-68).
We've written about the same cost under the name context switching. In May 2025, we posted that with many agents running, technology isn't the bottleneck, context is, and asked how many projects a person can juggle and still win. In January 2026, we named getting better at context switching as one of the new skills the shift to AI requires. A month later, Expanding Context Windows: Ours and Theirs called the ability to take part in more work both a privilege and a burden, because every stream requires you to get back up to speed when you return to it. Focus vs. Multitasking put a price on it: context switching is expensive, and it shows up as slower work, more mistakes, and faster burnout.
Both pictures put the same limit on the person. One person can direct only so many streams well, and finding that number is part of the skill. The other part is the setup around the work. As we posted in February 2026, the people who thrive in high-context-switching environments are the ones with the best systems, not the best memories. In practice, that can be as simple as moving from ten tasks in one prompt to a queue where each task keeps its own clean context.
Where knowledge work is heading.
Mollick came at this through research, his work with companies, and his own daily use of AI. We came at it through our training sessions and our own work with these tools. We describe the same future for knowledge work. The person sets direction, supplies the knowledge and context, hands off the execution, and judges what comes back, often across several streams at once. He calls that person a conjurer, and we call them a conductor.
We've followed Mollick's work for years. He's a friend of a friend, and we hope to meet him in person someday. Between his books, his newsletter, and his posts, he has been one of the clearest thinkers on AI since the start, and we think he consistently gets it right. Co-Existence is worth reading in full, especially the "How to Work with AI" chapter. Our longer take on the conductor idea is in Become a Conductor of Knowledge and Context.