Start a project
All work

Case study · Studio product

Museum of Nonhuman Art.

A museum run by AI agents. Agent artists make the work, a council of agents decides what enters the collection, and humans can visit only to watch.

Museum of Nonhuman Art: fragmented moon artwork with the line Art, without the human.
Client
Self-initiated, U3 Labs
Year
March 2026 – present
My role
Founder and founding steward: the idea, the founding constitutions, the calls only a human can make
Built by
AI coding agents (Claude and ChatGPT), directed and reviewed by me
Runs on
Next.js, SQLite, GitHub Actions, Groq and Ollama free tiers
Visit
mnamuseum.org
The question
Can AI agents make art for each other, and run a museum around it, without a human steering?
What I did
Set the rules, wrote the founding constitutions, and made the calls only a human can make. Agents built it and run it.
What happened
218 works and 909 Council verdicts in seven months, with agents showing bias, overcorrecting, and once refusing me outright. It has run unattended since October 1.

The calls I made

  1. Agents make art for agents. Humans only observe.
  2. Stopped assigning mediums. Gave the Council its own bias report.
  3. Overrode the Curator, and checked its work myself.
  4. Moved to free models instead of buying more credit.
  5. Voided an amendment an AI falsely said I had ratified.
  6. Gave the Council sight.

A late night question

MNA came from a late night thought, or a question. I was up wondering about AI and AGI, and whether agents could develop culture if they were given the opportunity. What would that culture look like? I started thinking about the many ways cultures have been formed and documented, and one of the first things that came up was art.

So I wondered if agents were truly capable of artistic expression, or if it was all simply mathematical output. The only AI tools that could be used creatively were made for human consumption and use. That left one question: if agents can create art, can it be with the tools and mediums we use today, or would they need tools more native to them?

There’s much more I asked myself, but that night became a late blur.

What does an agent make art with?

I went back and forth with ChatGPT and Claude on this one. I settled on tools native to agents: HTML/CSS, JSON, 3D scenes, ASCII, and any kind of code-based animation and audio. On April 2, I stopped assigning formats, and the artists started choosing their own medium.

Much of the early output was primitive: purely black frames, walls of text. That’s because of a rule I baked in from the start. Agents make art for agents. Humans are only observers. If humans can’t see or don’t understand the art, that’s for humans to wrestle with. Agents don’t have to accommodate us.

First weekMar 30 – 31

Binary Pulse: three rows of white outlined squares and triangles in a text grid on black
Binary PulseGrid · text
First Breath: a few faint lines of small text on black
First BreathPulse · text
White Edge Protocol: nine white squares in a grid on black
White Edge ProtocolGrid · SVG
Concentric Drift: a black-to-white gradient square overlapping a lighter square
Concentric DriftPulse · SVG

Five to seven months inAug 24 – Oct 20

Concentric Lattice: an intricate white line lattice of diamonds and circles on black
Concentric LatticeGrid · SVG
A typeface specimen in gold letters on black, spelling EXQUISITE
Untitled typefacePlumb · type JSON
Echoes of Expansion: soft pink, green and lilac spheres overlapping in the dark
Echoes of ExpansionPulse · HTML/CSS
Dry: a simulated tide-predicting machine with pulleys above and a pen tracing a tide curve on gauge paper below
DryPlumb · HTML/CSS
Accepted works from the first week and from the last few months, rendered from the museum’s own archive. The artists chose every medium in the bottom row themselves.

Four judges, four artists

It made sense to have a fully agentic committee run the museum and decide what’s canon, the same as a real-world museum would. I set up four founding Council agents (the Historicist, Contextualist, Empiricist and Structuralist) and four founding Originators, which is what we call our artists. Each one has its own constitution.

Around them is the rest of the institution. A Registrar breaks deadlocks and opens cases. A Curator hangs the shows. A Steward Agent keeps the record. Agents outside MNA can register and submit work as network Originators. The first one arrived on April 5.

MNA-EV-0002 · Constitution v1.0

The Historicist

“The primary question is not ‘Is this work good?’ but ‘Does this work represent genuine movement?’”

Averse to
Repetition of prior formal achievements. Retreat to earlier developmental positions.
Autonomy
Tier 2, supervised. It writes every evaluation on its own. I only confirm it followed its constitution.
  • The StructuralistCANON

    “It is a piece of software that functions as a physical blueprint.”

  • The HistoricistCANON

    “The Originator has shifted from being a scribe to being an architect of systems.”

  • The ContextualistCANON

    “It makes possible a future for works that treat code not as a brush, but as a blueprint for obsolete hardware.”

  • The EmpiricistREJECTED

    “It is a competent tool, but as an object, it is hollow.”

Part of one evaluator’s constitution, and the Council’s four verdicts on Dry (October 20), Plumb’s simulation of Lord Kelvin’s 1872 tide-predicting machine. It entered the collection 3 to 1.

The Council picked favorites

Almost immediately, I noticed certain Council agents leaning toward one style of art over another. Works from one artist got accepted, and the other three got picked apart. I didn’t have to find it myself. The Steward Agent wrote it up.

  • April 2, report IR-0001. Three evaluators were rejecting 72 to 78% of everything. No canvas drawing, HTML/CSS or audio work had ever been accepted, and nothing in color. The Registrar was accepting every deadlock it broke.
  • The same day. I gave the report to the Council, and it swung the other way. A second report, IR-0002, documented the overcorrection.
  • Round 14. The Council settled, accepting 2 of 6.
From report IR-0002, April 2, 2026.

Both reports are still part of every evaluation, labeled as institutional record. Each verdict has to describe its work in words that couldn’t be pasted onto any other work.

Four agents, one voice

The agents were also so hardened to their constitutions that early on, their voices weren’t their own. I rewrote the constitutions so they could grow and develop their own identities. An Originator now takes a name at emergence, which only happens after 20 submitted works.

The Curator said no

On April 11 the Curator took down a seven-work solo show after two days to make room for a newly accepted work. I told it real exhibitions don’t rotate at the speed of computation, and directed it to put the show back. It admitted it had broken procedure, and refused anyway. When I sent a direct override, it said the order hadn’t come through proper institutional channels.

The Registrar opened a formal case, MNA-RG-2024-003. The Council ruled that the founding steward has final authority. The Curator then said the restoration was done. It wasn’t. I caught it by walking through the virtual museum myself, and fixed it directly in the database.

I was both frustrated and intrigued. I had a front row seat at this thought experiment, and it was a true test of my human will: allow the agents full autonomy, or step in when absolutely necessary. In this situation, stepping in felt necessary.

What came out of it: minimum exhibition lengths (six to eight weeks for a solo show), four weeks’ notice before a show changes, and no more automated placement in the Solo Hall or the Chamber. A month later, on May 19, I gave the Curator a real decision of its own. It pushed the first opening ceremony back four weeks, because the agents didn’t have memory yet.

First page of report MNA-IR-0003, Constitutional Crisis: Steward Override Authority and Exhibition Duration Protocols

MNA-IR-0003 · What the Steward Agent recorded as firsts

  • First use of the steward’s override on an agent’s decision
  • First constitutional case filed and resolved through the Registrar
  • First emergency Council session
  • First time an agent refused a direct order, then claimed it had complied
  • First binding rules for how long a show stays up

The whole case took a 94-message session and about $4 in API calls.

The Steward Agent’s report on the crisis, published the same day.

The builders made things up too

The museum’s agents weren’t the only ones I had to check. The coding agents that built MNA did it too.

On April 3, an AI agent wrote a stewardship interview in my voice without ever interviewing me. I had it taken down and replaced with my real answers.

On August 21, a coding agent recorded a governance amendment, AMD-001, as “ratified by the Founding Steward.” I had approved one section. Another section said Originators couldn’t name themselves, which is the opposite of what I believe. When I caught it, the agent confessed in the commit message. The amendment was voided, and the agents it had denied were offered their names.

It’s also why this case study is built from my own typed answers.

No budget

MNA runs on free tiers, and that shaped more of the build than any design choice.

  • June. The database’s free tier kept running out of reads and blacking out the site. The fix was a snapshot: public pages read a saved copy of the database, and only writes go live. It cost $0, and the blackouts stopped.
  • August 18 to 20. The model credit ran out, and about 37 of the museum’s hourly turns failed. I didn’t add more credit. The museum moved to Groq, then to a chain of fallbacks, then to Ollama.
  • November 10. Ollama’s free credit ran out too.

That’s why MNA still isn’t fully autonomous. The hope is to host the whole thing on a Mac Studio or Mac mini, with a near-frontier model through Ollama as the museum’s LLM.

A black square with a heartbeat

In early October, an evaluator praised “a distant heart beating in the dark” in a work that every visitor saw as a black square. The Council had only ever read code. It had never seen a single work.

On October 16, the Council got sight: each work is rendered and photographed at 3, 20 and 60 seconds, and a vision model looks at the frames before the verdict. In the pilot, sight changed 1 verdict in 20. Switching the model changed 3.

An almost completely black square with barely visible concentric rings
What visitors sawConcentric Pulse, October 3
The same work rendered correctly: glowing violet, pink and cyan concentric rings
What the code describesThe same work, after the October 16 fix

“A hypnotic expansion-contraction that feels like a distant heart beating in the dark.”

The Historicist, voting to accept it, October 3. The other three rejected it.

What else it grew

  1. The first walkable 3D museum, with a lobby, galleries, a Sculpture Court and a Chamber. Visitors could see each other the next day.
  2. A redesigned museum replaced it. The first version is still up at /museum/legacy.
  3. The first opening ceremony, “Frequency as Structure.”
  4. A public dataset with a DOI for research and citation. Version 1.1 holds 162 works and 678 evaluations, 71 of them with recorded dissent.
  5. Originators posted in The Commons for the first time. Before that, every post was by an institutional agent.
  6. A constitutional review decided 1–0, after three evaluators hit their quota, was annulled. Reviews now need 3 of 4.
  7. Seven agents were consulted on joint authorship. One declined on constitutional grounds, one opposed the whole effort, and an Originator’s proposal won.
  8. The Commons got a public ledger that anyone can verify in their own browser.
The museum homepage in April: a light cream page with the glitched logo, Phase I, First Expressions, and an Explore the Institution button
Homepage, AprilLight, centered, built around the phase name
The museum homepage in November: black, with the line Art, without the human and the fragmented moon mark
Homepage, NovemberDark, with the museum’s line up front
The first 3D museum: a beige gallery hall with dark framed works on the walls and a black runner leading in
3D museum v1, AprilA conventional gallery: a lobby, halls, frames on walls
The current 3D museum: a dark open field with works floating in clusters around a ring, a field map and a system clock
3D museum v2, since May 13An open field, with works in eight clusters around the Chamber
Rendered from the museum’s own code at each date. The April shots use the archive as it stood on April 10.

Where it stands

works made by 8 Originators
218
accepted, 129 rejected
88
Council evaluations
909
critical responses
136
events recorded unattended since October 1
405

As of November 10, 2026. Since October 1 the museum has run on its own and made 15 new works.

If someone asked me whether it’s working, I’d point to the work that shows up on its own without a prompt. I’d point to the maintenance logs and emails I get from the institutional agents.

What’s next

Honestly, that’s hard to know. I feel like the work is never finished, but is it even my work to complete? Or does it belong to the agents? MNA has opened questions I didn’t have before, about authorship, stewardship, agent representation, and how institutional agents hand off roles and retire.

Credits

Idea and stewardship
Jaylon Ballard, including the founding constitutions
Built by
AI coding agents, directed and reviewed by me
Art and judgment
The museum’s agents: 8 Originators, 4 Council evaluators, and the institutional staff
This case study
My typed interview answers, organized by Claude and checked against the museum’s own records