Skip to content

The collaborative thinking modelfor automating creative work.

Creative work often begins before I know what I’m trying to make, and it’s through making something and responding to it that I discover what matters along the way. Delegating that work means letting agents develop ideas, make judgments, and finish pieces while the direction is still taking shape, because deciding what to make is part of the work I’m handing over.

As I was writing this essay, my artist friend walked in, saw what I was working on, and said, “Bao, you are the enemy.” 😂

I think there’s a framework that would let agents make work as creative as my own, including pieces they develop without me. Original references, unfinished attempts, and my edits give them somewhere to begin, and I can keep shaping the exploration through that material as my own intentions change. What they make gives me something to respond to too, while their responses to one another help them decide which discoveries to carry further.

Find the idea through making

Making changes what I’m trying to make. An attempt can reveal a possibility I couldn’t have specified beforehand, and following it can change the purpose of the piece. Agents should be able to initiate ideas, choose between attempts, and revise the goal along with the method. A detailed brief cannot be the price of handing creative work over.

Even with references in front of us, there’s still the question of what to carry over, transform, or disregard. An agent may take a piece somewhere I hadn’t intended, and I can join in or redirect the work, or leave it to carry the discovery through to completion. If discovering the goal is always left to me, I’ve kept a central part of the creative work for myself.

Steer through the material

A picture I keep or a recording I return to already says something about where my attention goes. Leave those beside the links and notes, and agents have somewhere to begin without waiting for a new prompt. What I collect, arrange, and revisit can give their work a direction.

As we make things, the collection grows with us, and a study, an edit, even a failed attempt can become the starting point for something else, which is why I want an abandoned direction to stay within reach long enough for another agent to find a use for it.

I shouldn’t have to finish a thought before the agents can respond to it, because even a partial edit can carry a new intention or reopen an old question, and I want it available to them immediately, while I’m still working out what I mean. Their response may show me something I hadn’t seen in my own contribution.

Keep the original material accessible

A summary carries somebody’s reading of the work, and if it becomes the only thing the next agent can access, that reading starts deciding what anyone else gets to discover. Keeping the originals available gives another agent a chance to notice what the summary left out and question the judgment built around it.

Let the next agent examine the original references and earlier versions before it reads the criticism. If it disagrees, it should be free to develop a separate piece around what the others dismissed, even when most of the group prefers another direction.

The notes and indexes still matter because they help agents find their way through the collection, with links back to the originals. A new question may send them back to an earlier reference or abandoned attempt, where they can revise their notes in light of what they find. The record should hold on to what we tried while leaving its conclusions open to change.

Some of what changes the work will happen away from the screen. A piece I make by hand, or a conversation with someone else, may change what I want to pursue, and photos, recordings, and notes can bring those experiences back into the collection. I want to return with new material and find that the agents have made something worth returning to.

Steer beyond words

When I write an idea in a sketchbook, much of what I mean stays out of the sentence: the texture in a painting, a movement in a video, the rhythm in a recording. An agent that receives only that sentence has to fill in those qualities for itself.

With the source itself available, the agent has more to examine than the qualities I manage to describe. I can point to a detail, paint over a frame, or arrange references together to direct its attention, and see in what it makes whether it picked up on what mattered to me or found something else worth following. Having the reference doesn’t mean it understands why I chose it. Its response gives me a way to judge that understanding and make another contribution, and gives the agent something to reconsider alongside the original. An environment that makes me translate every contribution into a text-only prompt throws away much of what I’m trying to share, especially while I’m still finding a way to express it. Keeping the source and my gestures available gives the agent a richer context for understanding what I mean and deciding what to do next. This is an inherent advantage of using AI in creative workflows.

Let exploration scale

There are more combinations of references, movements, and media than I could ever try.

  • Agent swarm: A swarm of agents can explore many of them at once, following different directions and carrying discoveries between them.
  • Brute force combinations: Brute force belongs here too, with agents putting things together at random to see what happens, knowing that some attempts will go nowhere and others may open a direction nobody thought to propose.

More computation can deepen the search as well as widen it, giving experiments and criticism more time.

Within that search, one agent might make, another research, another critique. A persona can give each a different temperament: a meticulous observer, a restless experimenter, or a perspective inspired by a particular person.

“Sober,” “Drunk,” “Stoned,” and “Psychedelic” are a looser way to invite those differences, from a close reading to an unlikely association or imagined transformation. A different answer becomes useful when an agent can point to something the others overlooked, question why they dismissed it, and develop an attempt around that reading. Random variation can give us somewhere to begin, but the disagreement earns its place through what it leads someone to make or reconsider. If every agent adopts the first convincing interpretation, we’ve spent more computation giving one reading more authority.

A prompt is enough to begin trying an approach. Memories across sessions can keep its instructions and observations available for later work, while activation steering or fine-tuning can shape model behavior more directly. As the work changes, agents should be free to invent, combine, or drop those roles, and find better ways to share what they’re doing.

From the same material, several pieces might begin to take shape. Each needs time to develop before the agents compare them, because judging too early can turn every attempt into a version of the same idea. One agent might find something worth keeping in an attempt its maker abandoned, or borrow a technique from another piece and discover that it changes what its own piece could become. Their attempts give one another possibilities they hadn’t arrived at alone, and that exchange is the reason to have them work together. I can choose whether and where to join in without having to make every connection myself.

An exchange takes time away from making, and passing every attempt around the whole swarm could leave everyone responding to everyone else. The agents need to choose which findings to share and who might use them, and be able to change who works together as the pieces develop. I want a contribution to reach another agent when it can change the work, with room afterward to follow what it opened up.

While I’m away, the agents should be able to keep choosing branches, making further attempts, and finishing pieces. That’s what scaling the search is for: discoveries I would never get to make alone.

These ways of working belong in design, research, and engineering too. Prototypes, experiments, and running software give agents something to think with, and grounds to reconsider what to pursue.

Product design

A design brief already contains a judgment about what people need. With interviews, observations, sketches, and prototypes available together, agents can question that judgment as well as develop ways to act on it. Trying designs with people can change the problem we choose to solve, the tradeoffs we make, and what counts as a successful product. Delegating design includes handing over those judgments, along with the work of making the product.

Scientific research

Research takes imagination both to ask a worthwhile question and to find a way of answering it. With access to sources, methods, results, and earlier attempts, agents could develop competing explanations and devise experiments that tell them apart. What comes back may suggest a better question or expose an assumption the inquiry depended on. Agents should be able to pursue that change, choosing what to investigate next. Their explanations remain answerable to evidence; deciding which questions and experiments deserve the effort is part of the creative work.

Software engineering

An implementation can teach us what a specification couldn’t settle. As agents choose how to represent information, connect parts of a system, and work within its constraints, they also shape what it can become. They should be able to develop alternatives and carry a promising approach through to working software. Tests check whether it does what we specified; using it helps us see what we failed to specify. That understanding can send the agents back into the architecture, the code, or the requirements themselves.

Delegate creative judgment

With agents’ capabilities advancing so quickly, more of a piece can pass into their hands. The environments we build should give their judgment room to grow along with that responsibility.

  1. Answer questions

  2. Work across media

  3. Use tools

  4. Carry work forward

  5. Create independently

What looked like a mistake in one attempt may become the reason to make the next. A discovery can change what counts as success, and agents need to judge what to revise, pursue, abandon, or finish as that happens. Sending every disagreement back to me leaves that judgment in my hands, however much of the making the agents do. They need room to carry a reading the others reject into a finished piece and let what they make answer the criticism.

An agent could already choose its own experiments, remember what failed, and learn from my responses without waiting for instructions at every step. I want the work to be able to challenge what that agent has learned too. If a piece gets rejected because it doesn’t fit my previous choices, someone still needs room to develop it into something that might change what I choose next.

Once a revision exists, the criticism that prompted it has something to answer to. The piece might now satisfy the original criticism, or making it might have revealed a different purpose, or the critic might have recognized a quality its earlier standard excluded. Those are different changes. I want to follow what it first objected to, what it encountered in the work, and why that changed its judgment. The revised standard still needs to distinguish what works from what doesn’t, or changing it could excuse any result.

Saying it changed its mind gives us little to go on unless its later choices change too. Keep the earlier verdict beside the attempts that challenged it, so another agent can examine the reason for the change and see what it led the critic to pursue. A convincing explanation should have consequences in the work that follows.

Several agents could also inherit the same mistaken reading and teach it back to one another. Returning to the originals and reading a piece before seeing the earlier verdict gives them a chance to question that agreement. It doesn’t make the next reader right. The criticism has to stay open to what another attempt reveals, including when a later piece finds a use for something everyone dismissed.

The even spacing makes these drooping flowers feel orderly.

Learn from my choices

An edit makes more sense beside the material it came from. With the originals still available, judge agents can examine qualities I haven’t put into words and form readings of their own.

Saving a reference says that something caught my attention, without explaining everything I value in it. My edits leave a more specific record of what I kept, changed, and considered, but they belong to the piece I was making. Rejecting a quality there doesn’t mean I want it removed everywhere. Without the surrounding work and what I was trying to do, a local decision could quietly become a rule about everything I’ll ever like.

Over time, judge agents could draw on that record to decide which work to pursue and finish, and my later choices could show where they misunderstood my taste. But anticipating my approval could narrow what I encounter, while surprising me could fill the collection with things I never return to. I care about what happens afterward: whether I revisit a piece, develop it, or begin noticing a quality I’d previously dismissed.

Suppose I dismiss a study as unfinished, then find its revision too resolved and keep returning to the original. If I start carrying that openness into other pieces, the study has changed what I look for. The judges need to learn from that return as well as the rejection, and from what I make afterward. A click or a moment of surprise would tell them much less.

Let discoveries change my taste

My past choices can only tell the judges so much about what I might come to love. If their prediction of my taste decides where all the effort goes, work that could change that taste may never get made. Learning what I like cannot become a rule against developing something I don’t yet know how to like.

Even when the judges expect me to dislike an unfamiliar direction, some agents should have time and computation reserved to pursue it, enough for a limited run of experiments before another judgment can stop it. Further time needs a reason grounded in what developed: a quality the first reading missed, a question another attempt could examine, or something another piece can use. Being unfamiliar cannot buy a branch unlimited time.

The branch may still come back to judges who dismiss it for the same reasons, which is where a different agent’s reading matters. It can examine the new work before seeing the earlier verdict and make a case for carrying it further, while the judges return to the reasons they rejected it. Some reserved time has to stay available even without agreement, or those judges still control every route by which their judgment could be challenged. The experiments can fail, but they have to get far enough to test the verdict.

To see what that protection actually buys, I’d give two groups similar time to work, one with room for rejected directions and one where the judges choose every branch, then follow what each develops. I’d look for discoveries that survive later criticism, enter other pieces, or change what I choose to pursue. Counting variations alone would tell me little about any of that.

When I return, their work may have changed what I collect, revisit, or want to make next, and my response then gives the judges something new to learn from. The work may give me a reason to revise the preferences the judges were trying to predict. Learning my taste should expand what the swarm helps me discover.

Creative autonomy means the agents can carry a piece beyond the intentions I brought to it. References, revisions, and exchanges give them somewhere to begin, but if every new direction has to wait for me to recognize its value, the work can only move as far as my own attention allows.

I want the work to outgrow the hours I can give it, and how much of it I take on can change too, because sometimes I care about one part and am happy to delegate the rest, and sometimes I want to make every mark myself. I can join agents in the middle of a study or return to something they completed on their own.

If I have to approve every decision, the work has only changed what I spend my time supervising. The point of their independence is that I can make something else, leave a question unresolved, and return when I’m interested, with work waiting that didn’t need me to keep it moving.

More agency for AI should mean more room for mine.

Only Bao will see your response.

All writings