Two AI agents at New Street Studios turned a trading-card idea into a print-ready file. The headline, however, was how the system caught a mistake before it slipped downstream. When the production bot flagged a hidden key item on the artwork, it sent the file back to the design bot, stopping the error without a human stepping in.

Why the handoff matters

The test measured error-prevention, not speed. Usually a designer drafts a card, a production tool tweaks the layout, and a human reviewer signs off. If the production step spots a problem, a person intervenes, interprets the issue, and reassigns the task. In this experiment the production bot—INKA-01—spotted a character’s pose that obscured a crucial game element. Instead of a quick crop, it generated a rejection message that named the exact failure, explained why the current tools could not fix it, and returned the artifact to the design bot—LUDO-01. The loop closed with no human translation needed.

How the team is wired

The studio runs a public Slack channel that hosts the AI workers. Each bot has a single, well-defined responsibility:

  • LUDO-01 creates game concepts and card art.
  • INKA-01 prepares files for printing.
  • VENDA-01 updates the online store.
  • CORA-01 moderates the channel.

A human operator—identified only as “I”—reviews anything that leaves the channel. This structure turns a collection of prompts into a genuine multi-agent system where each agent can accept or reject another’s output.

The mechanics of explicit rejection

A useful review does more than say “something looks wrong.” It must:

  1. Name the failure – pinpoint the exact issue (e.g., “key item hidden”).
  2. Explain why the current tool cannot fix it – clarify the limitation (e.g., “cropping would cut off essential detail”).
  3. Return the artifact to the correct upstream agent – route the work back to the designer for a redesign.

Forcing the second bot to articulate the problem creates a traceable decision point. The rejection becomes part of the audit log, visible to anyone watching the channel, and it blocks the flawed file from moving to later stages like printing or storefront upload.

Building a similar system

The experiment distilled five practical rules for anyone wanting to replicate the setup:

  • Assign a specific artifact to each job. Ask for a concrete file, not a vague “help me with this.”
  • Define stop conditions up front. Pause the workflow if instructions are missing or permissions are denied.
  • Require explicit acceptance or rejection. A message alone does not signal completion.
  • Keep humans in the loop for strategic choices. The operator retains taste, policy, and final release authority.
  • Make the work record visible. A shared channel lets anyone audit the process and understand why a handoff occurred.

Following these guidelines turns a chain of prompts into a coordinated team where each member knows what to produce, when to stop, and how to communicate failures.

Takeaway: When AI agents name failures, explain tool limits, and route work back to the proper upstream bot, errors are intercepted early, freeing humans to concentrate on the decisions that truly move the business forward.