Inspiration
I'm a creator myself — I run a faceless YouTube channel about mythical creatures. The least creative part of the job eats the most time: hours of research across trend reports, competitor uploads and audience threads before a single word of script exists. That pre-production grind is the "critical bottleneck" of the media workflow — so I built the agent that kills it.
What it does
Describe your channel in one sentence — any language. ShowScout:
- Plans and runs 2–4 live web research calls through Parallel's Search API with different objectives: niche trends, competitor output, audience pain points.
- From evidence only — never invented — selects the 5 strongest video topics. Each comes with a grounded "why now" reason, a one-line hook, and the strongest source URL.
- Writes a beat-by-beat script outline for the #1 topic: cold-open hook, beats, payoff, CTA.
- Saves the slate to Firestore — a real, persistent artifact the creator can come back to, not just chat text.
- Replies in the creator's language, with the full tool-call log visible in the UI.
My favorite run: asked about my own mythology niche, the agent discovered that the biggest faceless channel in the space (1.6M subs) stopped uploading in 2025 — and built the whole slate around filling that gap. That's a real strategic insight I didn't have.
How we built it
- Google ADK 2.6 — one agent, three plain-Python tools (
web_research,save_content_slate,list_slates). - Gemini 3.5 Flash via Vertex AI — planning, evidence selection, multilingual replies. ADC on Cloud Run, zero API keys in the container.
- Parallel Search API (
parallel-webSDK) — every research step hits the live web at runtime with objective + queries; results come back as excerpts with URLs and publish dates. - Cloud Run (asia-southeast1) hosts the FastAPI app + showrunner UI, deployed from source.
- Firestore persists slates across instances and restarts.
Challenges we ran into
- Agent-shaped API usage: one run = multiple LLM turns + several Parallel calls. Free-tier Gemini API quotas died instantly; Vertex AI with ADC was the right architecture.
- Keeping topics honest: the instruction forces "evidence only" and every topic must carry a source_url from the research — the model can't invent trends.
- IAM propagation on fresh service accounts taught me to wait before declaring things broken.
Accomplishments we're proud of
A creator in any niche gets a research-grounded, production-ready slate in about a minute — work that used to take an afternoon. And the agent's insights are real: gaps in the market, competitor moves, seasonal angles, each with a link to check.
What we learned
Agents earn trust when the LLM plans and the tools decide. Parallel gives the agent real eyes on the live web; Gemini turns evidence into judgment; the code guarantees every claim has a source.
What's next
Channel-connected mode (pull the creator's actual upload history and analytics), scheduled weekly slates, one-click export to a script editor, and multi-property support for creator networks.
Log in or sign up for Devpost to join the conversation.