Unprompted
What happens when we give AI power to act on our behalf?
I often describe my feelings on AI as a mix of awe and existential crisis. When I work with the frontier models, I’m often struck by what they can deliver. Then, I immediately get this sinking feeling. Something like guilt, or like I stole something.[1] I don’t think I’m alone in feeling this way, but solidarity is limited comfort.
A quick note before we go further: agentsandagency.com has grown well beyond the blog.[2] I think it is important to actually use AI products if you are going to write about them, so I’ve deliberately taken a step back from writing to make sure I’m not just working in the abstract. That building is also partially responsible for why I no longer believe it is appropriate to call AI a “tool.”
The Hammer and the Playhouse
I’m not usually a fan of analogies,[3] but humor me. If you primarily engage with large language models via a chat interface, they do seem mostly like technology tools. Summarize this email, turn these bullet points into a memo, tell me the key people on topic X. With this approach, the internet probably seems like a hammer and an LLM probably seems like a nail gun. But products like Claude Cowork, Claude Code, and OpenAI’s Codex are something different. Why are you hammering something in the first place? Is it to build a playhouse? Tell these agentic models your goal and they are more than a hammer; they are the general contractor looking up local ordinances, sourcing materials, designing the layout, and then, sure, hammering too. This is powerful stuff and it brings lots of concerns.[4] But I want to go a layer deeper.
What happens when AI agents know enough about you to suggest it’s time to build a playhouse without you asking? Once you feel comfortable with the performance of the agent, maybe you won’t even ask; it will just know you need one and build it. Obviously not a physical playhouse, since robots manipulating the physical world still lag way behind human capabilities.[5]
Is “On Our Behalf” in Our Best Interest?
Set the analogy aside for a moment, and return to what these AI “tools” can do. They can increasingly deliver complicated, long-form, multi-step work, and the length of task they can handle keeps doubling.[6] When these platforms get enough information about us, the next step will be creating on our behalf without us ever formally identifying the goal, task, or project. We will have things built for us unprompted.[7]
The earliest versions are already here. OpenAI’s Pulse does research overnight by drawing on your chats, your memory, your connected calendar and greets you each morning with cards you never asked for. OpenAI’s stated ambition is for progress to happen “even when you are not asking.”[8] The industry has started calling this category “ambient agents.” Whatever we call it, the direction is clear.
Right now, I could build an AI agent to read my emails, prepare responses, and send replies. I’m talking about the next level, where the agent has enough context about me to identify someone I don’t know that I should know, reach out to them on my behalf, and write to them in my voice. This goes beyond removing the friction of learning to something more like removing the friction of… being, I guess?
Will This Make Us Happy?
I’ve been struck by the findings in the Walton Family Foundation-Gallup research on Gen Z’s attitudes about AI.[9] To oversimplify a bit, there are three combined findings that may not be surprising, but are alarming: 1. Most report using AI — just over half now use it weekly or more. 2. Most have a negative view of AI along some dimension — anxiety holds at four in ten, anger is rising, and excitement is falling. 3. Few have sufficient training — in the companion survey of heartland Gen Zers, only about one in ten workers felt extremely prepared to use AI on the job. Collectively, that’s, uh, not great. And I don’t think this is just a Gen Z story. Gallup’s workplace research finds half of U.S. employees now use AI in some form, while only about a quarter say their employer has clearly communicated how it should be used.[10]
If AI agents start creating things for us that they think we will like, will these sentiments improve? I fear the opposite, because we will have given up our agency. We have run a version of this experiment before. Recommendation feeds already decide, unprompted, what billions of people see every day, and few would say the arrangement made us calmer or more in control. Agentic AI extends that logic from selecting our content to producing our work, our words, and maybe our relationships.
One of the original motivations for this blog[11] was to focus on how we maintain our dignity through human agency in a world that is going to be increasingly full of AI agents. The convenience of AI responses and products requires deliberate effort on our part to confront. The old adage to “think for yourself” may be cliché, but it’s a necessary first step.
Maybe that’s also what the guilt at the top of this post is about. The sinking feeling shows up when something gets made and I wasn’t fully in it. As AI agents become increasingly capable and gain enough context to generate on our behalf, it will take even more deliberate effort on our part to “be human.” Here’s one place to start: remember that “unprompted” used to be a human word. It’s the kid who cleans her room without being asked, the friend who calls out of nowhere. The machines are learning to act unprompted. Don’t hand over the world without a fight. Want things no algorithm suggested, reach out to people no agent picked for you, build the playhouse nobody told you to build.
[1]Given how the models are trained, I am always concerned that parts of real humans’ IP have been stolen and given to me, but I don’t think this explains everything for me.
[2]The site now includes more about my journey and work on AI, plus an area called “the lab” where I am building prototypes of what is possible with these systems.
[3]Yes, technically the fallacy is the false analogy, not analogy itself. But arguments by analogy break easily, which is exactly why I distrust them — including my own.
[4]See my previous post on the liquidity crisis of expertise: https://agentsandagency.substack.com/p/the-liquidity-crisis-of-expertise
[5]When robots have their “ChatGPT, November 2022” moment, things are going to get really weird. Roboticists even have a name for why it is taking longer: Moravec’s paradox: the “hard” cognitive problems turned out to be easy for machines, and the “easy” physical ones turned out to be hard.
[6]METR estimates the length of tasks frontier AI agents can complete (at a 50% success rate) has been doubling roughly every seven months since 2019, and faster since 2024: https://metr.org/time-horizons/
[7]Oh exciting, it’s like when a character in a movie says the name of the movie in dialogue!
[8]OpenAI, “Introducing ChatGPT Pulse” (September 2025): https://openai.com/index/introducing-chatgpt-pulse/
[9]The 2026 installment, “The AI Paradox,” is worth a read: https://www.gallup.com/analytics/708176/gallup-walton-voices-gen-paradox.aspx. The companion survey of heartland Gen Zers, conducted with Heartland Forward, is here: https://news.gallup.com/poll/660302/heartland-gen-zers-feel-unprepared-work.aspx
[10]Gallup’s workplace AI tracking (2026): https://www.gallup.com/workplace/704252/workplace-separates-adopters-holdouts.aspx
[11]And literally why it is named “Agents and Agency”!