FAUNA: A creative agent built for visual exploration

FLORA began as a node-based canvas designed to help creative teams explore visual directions for fashion campaigns. For power users, the canvas is a flexible tool: it lets you lay out steps, branches, and detailed prompts to control every piece of an image generation.

That power came with friction. Every creative had to think like a workflow engineer, configuring prompts and model choices for each node before they could start exploring. The setup interrupted the creative process rather than supporting it.

FAUNA was built to remove that burden. Instead of starting from a blank canvas, users describe what they want—a campaign visual, a moodboard, or a lookbook direction—and the agent pulls references, selects models, and generates variations for the user to refine. Alec Jo, Head of Applied AI at FLORA, describes the goal this way: “Unlike agents that assume that there’s one answer, ideating is inherent to the creative process. We built FAUNA to be a creative partner that explores and iterates with you, rather than trying to give you the answer.”

Why visual agents need durable orchestration

Text-based AI tasks can lull you into underestimating what visual work demands. A single image or video generation can take minutes, and a professional creative session fans out into many concurrent jobs: different angles, backgrounds, styling directions, and “what if” branches. Those jobs run in parallel, finish at unpredictable times, and still need to combine into a path the user can navigate.

Early on, FLORA’s team built on LangChain primitives and considered Temporal for orchestration. The setup worked but split the system in two: separate maintenance, separate abstraction layers, and constant plumbing work just to keep the product moving.

A single AI stack, start to finish

The team migrated to the Vercel AI Stack to consolidate building and deployment. The key pieces:

  • AI SDK and its agent framework provide the primitives that power FAUNA’s behavior—models, tool calling, and the agent logic layer.
  • Workflow SDK’s DurableAgent adds durable orchestration to the AI SDK agent framework so each step in the loop persists, retries on failure, and runs indefinitely without losing state.
  • Fluid compute runs jobs cost-efficiently as the agent fans out into many concurrent tasks, which matters for parallel, long-running image generations.

For FLORA, the shift wasn't about swapping one tool for another. It was about moving to an integrated system where primitives, orchestration, and compute are designed to work together. As Jo puts it: “We stopped having infrastructure debates and started having product debates.”

Process as product

FLORA’s larger ambition is to support all design work that isn't UI/UX, and FAUNA represents their first step in orchestrating visual ideation end-to-end. Building the agent also changed how the team thinks about its users. In creative work, the process is the product, and maintaining flow state for longer leads to better outcomes.

“Questions like 'how do we make this durable' or 'what if the user disconnects mid-generation' used to eat up design discussions. Now they're just solved, so we have the freedom to ship as quickly as we experiment.”

That speed matters because new image models and capabilities arrive weekly. With infrastructure out of the way, FLORA can take advantage of them as quickly as they appear.