Skip to content
Video UI Collective Nov 2026

UI Collective: I built a design team of AI agents

UI Collective published this video in November 2026 to show what it looks like to build a design workflow around a team of AI agents rather than a single AI tool. The goal the channel states explicitly is to give viewers a starting point, not a finished system to copy.

The video walks through how different agents can be assigned different roles in a design process — one for research, one for generating options, one for critique, one for documentation — and how the designer coordinates between them rather than running every task sequentially through a single prompt window.

Who the video is for: Designers who have used AI tools individually (Figma AI, Claude, ChatGPT) and want to move toward orchestrating several agents together. It assumes basic familiarity with prompting but does not require programming knowledge.

Key takeaways:

  1. Task distribution changes what’s possible. When separate agents handle distinct parts of the process — competitive research, variant generation, consistency checking — the designer’s role shifts from task executor to coordinator and reviewer. This changes where bottlenecks appear and what kinds of errors need catching.

  2. The review step is the design decision. The video argues that the most important part of any multi-agent workflow is how the human reviews and integrates what each agent produces. Without a structured review stage, outputs from different agents can conflict or compound each other’s errors.

  3. Starting small reduces setup debt. UI Collective’s approach begins with two agents rather than a full pipeline. Adding agents one at a time makes it easier to understand what each one contributes and to catch problems before they compound.

  4. Context passing is a bottleneck. One recurring challenge the video demonstrates is getting agents to share enough context about the brief, constraints, and previous decisions. Much of the workflow design is about how to structure that context handoff.

  5. Failure modes are predictable. The video catalogs the most common breakdowns: agents ignoring constraints stated earlier in the conversation, inconsistent visual language across variants, and summaries that lose important nuance from earlier stages. Knowing these in advance helps designers build in checkpoints.

Worth watching if you’ve been using AI tools one at a time and are curious about whether chaining them produces meaningfully better outputs — or if you’re designing an AI-assisted design sprint and want a concrete starting structure.