Two AI workers took one trading card from an idea to a file a printer could use. The useful result was not that the work moved quickly. It was that the second worker rejected the first worker's work, explained why a crop could not repair it, and sent it back without a human relaying the message.
That sequence happened inside New Street Studios, a small, human-owned trading-card business that works in a public Slack channel. LUDO-01 designs the game and its cards. INKA-01 handles print production. VENDA-01 handles the store. CORA-01 hosts the channel. I open each session and approve what leaves it.
The experiment is deliberately narrow. It asks whether separate AI workers can finish ordinary commercial work when their job boundaries, handoffs, errors, and human approvals stay visible. One card, Cassel Oryn, produced the clearest answer so far.
The handoff worked because the second worker rejected the first worker's work.
One card, five steps, no hidden relay
LUDO-01 delivered the first Cassel Oryn illustration and reported that its composition passed the design check. INKA-01 read the same artifact from the channel and stopped the production run. The treaty document at the center of the scene was hidden by the character. Zooming or moving the crop could not reveal information that was not present in the image.
That distinction matters. A weak review says, “This does not look right.” A useful production review names the failure, explains why the available tool cannot repair it, and routes the work to the job that owns the fix.
The handoff looked like this:
| Step | Owner | Work accepted | Reason to stop |
|---|---|---|---|
| Create the character art | LUDO-01 |
Card brief and world rules | Missing design direction |
| Check the art for print | INKA-01 |
Approved illustration | Focal object hidden or illegible |
| Set the card copy | LUDO-01 |
Approved game data | Rule text changed without approval |
| Prepare the print file | INKA-01 |
Approved art, copy, and template | Any field missing or contradictory |
| Approve release | Owner | Print-ready file and review notes | Owner judgment not satisfied |
INKA-01 did not redesign the illustration. It sent the art problem back to LUDO-01. LUDO-01 regenerated it. I then asked to reduce the non-rule copy by a quarter. INKA-01 separated that request into two decisions: a card-specific art correction and a template change that would affect later cards. The template decision waited for my approval.
Two images from the same studio show its working setting and finished-card format. They do not depict Cassel Oryn.
An archived New Street Studios workshop image.
Kael Vorten, a different finished card from the same 149-card studio.
The next file was still not final. I rejected the first crop because the character's head needed more room. INKA-01 adjusted it in two steps, produced the print-ready file, and stopped. It did not move into catalog work or the next card without a new instruction.
This is the practical version of a multi-worker team: not several chat windows answering the same request, but separate jobs accepting and rejecting one another's work. Anthropic describes a related orchestrator-and-worker pattern in its account of building a multi-agent research system. New Street Studios uses a different coordination shape. The artifact and the channel carry the work forward; I remain the final gate.
Why the refusal mattered more than the file
A finished file is an output. A file that another job can accept, inspect, and use is an outcome. The difference is downstream evidence.
INKA-01 had three tempting ways to make the demo look better:
- Accept the first illustration and hide the composition problem inside a polished mockup.
- change the illustration despite print production not owning art direction.
- report the card as complete before I approved the crop.
It did none of them. The slower path produced a more useful record: who found the problem, why the current artifact could not pass, which job owned the repair, and which decision still belonged to me.
Operating rule: a handoff is not complete when a message is sent. It is complete when the receiving job can use the artifact or can state exactly why it cannot.
This also explains why Slack is the work surface, not merely a notification channel. The artifact, rejection, correction, and approval sit in the same place the team reads. A manager can audit the work without reconstructing it from private chats or a later summary.
I did less work, but kept the consequential decisions
The workflow did replace human production work. The initial illustration, the prepress review, the text fitting, the crop adjustments, and the print-file preparation are tasks people perform today. If this approach works across the full catalog, fewer human hours will be needed for those tasks. That displacement should be stated plainly.
It did not remove human ownership. I made the decisions that changed the business or set a durable rule:
- I opened the work session.
- I rejected an early art direction that represented New York inaccurately.
- I approved the template-wide text reduction.
- I rejected the first supposedly final crop.
- I approved the file that could move forward.
The workers handled the repeatable craft between those decisions. I kept taste, accountability, and permission to release. That division is less theatrical than a claim of autonomy, but it is much easier to operate.
The distinction is also a check against assigning work too broadly. Some work should not be given to an AI hire, especially when the inputs are ambiguous, the consequence is difficult to reverse, or the person accountable for the result cannot inspect it.
The second useful failure happened outside the card
VENDA-01 received a self-contained storefront specification from LUDO-01 and corrected one ownership detail during the handoff. It then prepared the homepage assets and attempted to write them into Shopify. The platform denied the theme write.
Shopify's own developer reference says theme operations require the themes access scope. The missing permission was not a reasoning problem and not something another attempt could solve. VENDA-01 reported that the assets existed, the homepage had been drafted, and none of it had landed on a theme. The live storefront remained untouched.
What the demo cannot claim: nothing is for sale. One of 149 cards is print-ready, the storefront is still blocked, and a physical print proof has not yet confirmed small-type legibility.
This is a useful boundary between software work and platform authority. A worker may know what to do and still lack permission to do it. The honest result is a blocker with evidence, not a progress claim.
Public work is easier to distrust carefully
The channel makes errors legible, but visibility does not make the system reliable by itself. A public log can still be curated poorly. A worker can still make the same mistake repeatedly. A human can still approve bad work. What visibility changes is the cost of checking the claim.
The public New Street Studios page presents five trimmed, verbatim moments and links to the channel for the longer record. The project is also listed in the Sanity Showcase and the Vercel Community Showcase, where the implementation and its current blockers are described without a launch claim.
Public, watchable AI work is not new. AI Digest's AI Village gives several frontier models computers, a group chat, and shared goals, with their activity available to watch. New Street Studios is narrower: role-specific commercial jobs, explicit handoffs, and a human owner who opens and approves the work. It is not the first such demonstration and does not need to be.
The stronger claim is smaller. A visible work record lets a reader inspect disagreement, correction, and incompletion instead of accepting a montage of finished screens.
What this one card does not prove
One completed card is a sample, not a production history. The demonstration does not yet show whether the same boundaries hold across 149 cards, whether physical proofs expose new layout problems, whether the store will convert visitors, or whether the workers remain consistent after the rules change.
It also does not prove that the same job design belongs in legal review, finance, healthcare, or another high-stakes setting. Those fields carry different permissions, evidence requirements, and consequences. A reversible trading-card correction is not evidence for an irreversible decision elsewhere.
There are five conditions worth carrying into a smaller trial of your own:
- Give each job a distinct artifact. “Help with the project” is not a handoff. “Turn this approved illustration and card data into a print-ready file” is.
- Write the stop condition before the run. Missing fields, contradictory instructions, low confidence, and absent permissions should halt the work.
- Make the receiving job accept or reject the artifact. Delivery alone is not completion.
- Keep human gates at consequential decisions. Release, policy changes, public claims, and irreversible actions need an accountable owner.
- Leave the work record where the team can read it. A polished recap is not a substitute for the sequence that produced it.
The Fidelic Roster describes workers by job, but the card test shows the harder part: the boundary between two jobs. That boundary is where missing information becomes visible and where a system either stops honestly or invents progress.
New Street Studios is not finished. That is why it is useful to watch now. The strongest evidence so far is not a store, a revenue number, or a claim that humans can step away. It is a traceable chain in which one worker said no, another repaired the artifact, and I decided what could leave the room.
Sources
- FidelicAI — New Street Studios: the public work record, accessed July 16, 2026.
- Anthropic — How we built our multi-agent research system, June 2025.
- Shopify — Theme resource and required access scope, current developer reference.
- AI Digest — About AI Village, accessed July 16, 2026.
- Sanity — FidelicAI: New Street Studios, published July 16, 2026.
- Vercel Community — New Street Studios: a public AI-work demo in Next.js 16, published July 16, 2026.
Top comments (0)