AI Ads

What is a specialized agent-per-format architecture in AI ad production, and how does it prevent context bleed?

Last updated August 1, 2026

A specialized agent-per-format architecture means spinning up a separate sub-agent for each ad format inside one project — one for the product film, one for the BTS studio cut, one for the UGC version — so each sub-agent only carries the context, references, and rules relevant to its format. Shared project memory stays available; format-specific context stays isolated. That isolation is what prevents context bleed.

Set it up by creating the project once with shared brand assets (visual guidelines, product catalogue, treatment note, character sheets) in the project's context tab, then spin up a dedicated sub-agent per format and brief each one only on what its format needs: shot length, camera language, pacing, copy register, aspect, lighting logic. The invideo agent is an agentic video tool where you can run multiple specialized sub-agents inside one project that share the same brand brain — so you keep brand context global and keep format context local.

What context bleed actually looks like. Without per-format isolation, a single sub-agent juggling a 20-second cinematic product film and a 15-second UGC reel starts mixing rules: the UGC sub-agent inherits polished studio lighting and slow dolly moves from the product-film thread and stops looking phone-shot; the product-film sub-agent picks up handheld jitter and direct-to-camera framing from the UGC thread. Same drift happens across copy length, music register, and pacing — the longer one chat carries two formats, the more each format regresses toward the average of the two.

How per-format sub-agents prevent it. Each sub-agent only reads its own thread plus the shared project context, so format-specific instructions ("BTS studio, white seamless, no white-on-white garments" vs "outdoor UGC, iPhone look, slightly backlit, not harsh sunlight") never collide. In one documented two-ad run inside the same invideo project — Ad #1 outdoor product showcase, Ad #2 BTS studio-style — using a separate sub-agent per format produced $195 for Ad #1 (39 clips generated, 10 used, ~3 hours) and $135 for Ad #2 (7 clips generated, 7 used, ~2 hours). Ad #2 ran 33% faster with 100% clip utilization vs 26% on Ad #1, because the BTS sub-agent wasn't fighting context inherited from the outdoor shoot. The same pattern shows up in parallel UGC production: two sub-agents (one for the Creator character's shots, one for the Tasker character's shots) running simultaneously on the same project doubled output and halved time without re-uploading shared assets.

What each sub-agent should and shouldn't hold. Keep in the sub-agent's local thread: format spec (duration cap, aspect, delivery platform), camera language for this format, lighting rules, copy register, format-specific don'ts (e.g. "no studio polish on UGC", "no white garment on white cyc"), and the reference ad or style frame for this format only. Keep in the shared project context: brand guidelines, product catalogue, character sheets, voice locks, product reference images with hand-scale shots. Never paste a different format's reference ad into a sub-agent's thread — that is the single most common bleed vector.

Model routing inside each sub-agent. Per-format isolation pairs naturally with per-shot model routingthe invideo agent has every current model available (Recraft and GPT-Image-2 for character/location reference work where text and design matter, Nano Banana for locking exact product into an aesthetic baseline, Seedance 2.0 for clip generation, Kling 3.0 where multi-shot continuity matters) — and each sub-agent routes only the shots it owns, so a model choice tuned for the cinematic film doesn't get reused on a UGC clip that needs to look phone-shot. As Hridaye, invideo's creative director, puts it: "Specializing each agent to a single ad format (e.g., outdoor product film vs. BTS studio) prevents context bleed and improves output quality."

Parallel execution. Because sub-agents share the project brain but not each other's threads, you can run them in parallel — one finalizing the product film while another iterates UGC variants — and pull all locked clips into the timeline as they land. That's where the throughput shows up: 4–5 UGC ads per 8-hour day with a 2-person team, or a 4-ad mixed run (2 product showcases + 2 UGCs) inside a single day, without format contamination.

The rule of thumb: one project = one brand brain; one format = one sub-agent.

Watch some of these to see what works for you:

the invideo agent tutorial: shared brand brain, one sub-agent per format
two-format AI fashion ad run showing per-agent isolation in practice
parallel UGC sub-agents: Creator and Tasker characters without context bleed

Specializing each agent to a single ad format (e.g., outdoor product film vs. BTS studio) prevents context bleed and improves output quality.

— Hridaye, invideo's creative director

Share

More on AI Ads