CineOS: Agentic Production Recovery

CineOS: Agentic Production Recovery

Film and television productions can lose expensive shoot days when an actor, location, or critical resource becomes unavailable at short notice. CineOS is an agentic production-recovery control room that turns that disruption into a clear, evidence-based recovery recommendation.

For the demo scenario, actor Maya Chen becomes unavailable for Scene 42 on the next production day. CineOS evaluates the incident through a coordinated team of specialist agents:

• Incident Intake extracts verified facts and unknowns.
• Script Intelligence retrieves the affected scene’s production constraints.
• Schedule Validation checks cast and location availability and validates a replacement-scene candidate.
• Budget Analysis compares schedule, cost, and creative-risk tradeoffs.
• Recovery Decision synthesizes the evidence into one safest recommended action.

The deployed workflow validates that swapping Scene 42 with Scene 47 is feasible: Scene 47 does not require Maya Chen, its required cast and location are available, and the change avoids a planned day of delay. CineOS clearly distinguishes verified facts, estimated impacts, recommended actions, and unresolved unknowns.

How we built it

CineOS is built with Python and Google ADK. A sequential multi-agent workflow uses Gemini on Vertex AI for reasoning, while deterministic production tools provide scene, availability, schedule-impact, and budget-impact evidence.

The project is deployed as a public Google Cloud Run service. Vertex AI credentials are provided through a dedicated Cloud Run service account, and the Grafana MCP credential is injected securely from Google Cloud Secret Manager rather than stored in source code.

Grafana is integrated through the Model Context Protocol (MCP) as an optional observability layer. The decision agent only reports Grafana data when it receives useful, real evidence; otherwise it explicitly states that observability evidence is unavailable instead of inventing telemetry.

Challenges and lessons

The key challenge was making the recommendation trustworthy rather than merely plausible. We designed strict agent boundaries so each specialist contributes only the evidence it owns, and the final agent cannot treat a low-cost option as automatically safe.

Deploying the ADK workflow to Cloud Run also required packaging the MCP dependency correctly and handling Grafana credentials with Secret Manager. The result is a reproducible cloud deployment that keeps secrets out of the repository while preserving the local development workflow.

What’s next

Future versions could connect CineOS to live production schedules, crew call-sheet systems, and richer Grafana dashboards, while preserving the same evidence-first and no-fabrication decision policy.

Built With

• cloud-run
• cloudbuild
• gemini
• google-adk
• google-cloud
• grafana
• mcp
• python
• secret-manager
• vertex-ai