Marketing technology · Creative platforms
An agent pipeline that generates brand-compliant marketing creative in the background, returns a session handle in under a second, and polls to completion instead of blocking the browser.
A B2B marketing-technology company needed a way for its customers to generate on-brand marketing creative at volume: a brief in, a finished, brand-compliant asset out, without a designer touching every iteration. The platform had to hold two things in tension — fast for the person waiting on it, and safe for a brand that cannot ship an off-guideline asset.
A multi-step agent pipeline generates a creative from a brief, scores it against brand and quality criteria, and loops until it clears both or hits an iteration cap. A separate ingestion path lets a brand team upload a guideline PDF or point at a live site; an extraction agent parses palette, typography, voice and logos into a template a human confirms before it becomes the brand's live standard, never automatically. Approved creatives can be refined, turned into channel variants, or extended into video from the same base.
Every AI action returns in well under a second with a session ID — the generation itself runs 45 to 90 seconds in the background — so the client survives a dropped connection or a mid-pipeline redeploy instead of hanging on an open request. Exactly one service writes to the relational database: agents receive full context up front and hand back results for the API to persist, so tenant isolation has exactly one enforcement point. Finished assets live in object storage behind short-lived signed URLs; the database never stores the pixels.
The three rules that make thirteen agents legible without a diagram: no endpoint waits on an agent, one service writes the database, and every result comes back through a declared, versioned key — never parsed from streamed text.
Generation, brand ingestion, the refine-approve-variant lifecycle and analysis-driven regeneration all run end to end in production today. One session type — translation into other languages — is wired into the routing and the agent service but does not yet produce a working result, and closing that gap is scoped as its own workstream. We treat "exists in the code" and "works for a user" as two different claims, and publish only the first as done when that's what's true.
Tell us what system the answer lives in and who needs it. We'll reply with a view on whether it's a two-week assessment, a five-week pilot, or something else.