Inspiration

Music-video production still feels like a relay race of briefs, treatments, shot lists, and renders—with almost no shared visibility into cost or failure. We wanted a director that thinks in departments (Writer, DP, Editor) and a control room that shows what those agents actually spend and break—powered by Grafana Cloud MCP + Agent Observability.

What it does

Creative Director AI takes a track brief and runs a music-video swarm:

  1. Resolves genre (hybrid Apple/Spotify-style catalog → ~25 deep packs or family templates)
  2. Delegates treatment to Writer, beat-synced storyboard + Veo prompts to DP
  3. After human approval, renders cuts (Veo), assembles (Pixels), and records provenance (C2PA)
  4. Uses Grafana MCP tools to investigate pipeline health, and Agent Observability + a live cost/spend dashboard to see generations, latency, and spend

How we built it

  • Google ADK (TypeScript) coordinator with A2A Agent Cards for Writer / DP / Editor
  • Genre system: static catalog + family templates + deep markdown visual bibles
  • Production tools: Vertex Veo, GCS, Pixels headless assembly, C2PA provenance
  • Grafana Cloud MCP (MCPToolset) for Loki/Tempo/Prometheus investigation
  • @grafana/agento11y + OTel OTLP into Grafana Cloud for generation traces
  • Local Dev UI via adk web, deploy path via Agent Runtime / agents-cli

Challenges we ran into

  • ADK Dev UI session not found after restart when App.name didn’t match the route app (agent)
  • Grafana Cloud MCP: no interactive OAuth in ADK Streamable HTTP → needed service-account tokens
  • Optional peer @modelcontextprotocol/sdk missing / unresolvable from ADK temp bundles → direct dep + NODE_PATH launcher
  • Genre alias over-matching (cinematic hip-hop → classical) → token/word-boundary matching
  • Keeping observability on without poisoning unit tests (dotenv override / skip flags)

Accomplishments that we're proud of

  • End-to-end A2A music-video swarm with HITL before expensive renders
  • Hybrid genre coverage beyond a 3-pack toy set (~25 deep packs + templates)
  • Real Grafana MCP + Agent Observability wiring—not just a dashboard screenshot
  • Live agent cost & spend overview on Grafana: https://thecreative.grafana.net/goto/s9bnpl
  • Soft-fail MCP so a bad observability hop doesn’t kill the creative turn

What we learned

Agent systems need the same ops discipline as production apps: auth modes, peer deps, session identity, and eval of “did the pack actually get passed?” Observability isn’t optional for multi-agent pipelines—it’s how you debug Veo/assembly failures and justify spend.

What's next for Creative Director AI

  • Live catalog sync from Spotify/Apple APIs
  • Stronger pack-handoff guarantees (not prompt-only)
  • More deep genre packs + eval harness for genre routing
  • Remote A2A specialists in Cloud Run
  • Deeper Grafana alerting on render failure rates and cost budgets

Built With

Share this project:

Updates

Submission history