Build AI Agents to Automate Film Production at Google's Agentic Cinema Hackathon

Google Cloud has opened submissions for Agentic Cinema: The Blockbuster Hackathon. The challenge asks developers to build AI agents that can handle real work in film and media production instead of just answering questions.

The official page lists a total prize pool of $75,000 split across five partner tracks. Each track (IBM, Grafana, Parallel, ClickHouse, and Replit) offers $7,500 for first place, $4,500 for second and $3,000 for third. Earlier announcements mentioned $60,000 while new partners were still being added, which is why that figure appears in many posts right now.

Submissions close on September 7, 2026. You pick one partner track and compete only against others in that same track. The project must use Gemini (compare Gemini 3.6 Flash vs 3.5 Flash) together with Google Cloud Agent Builder and connect to the chosen partner’s MCP server (learn more about MCP Server Workflows). The agent (or multi-agent setup) has to perform actual multi-step actions—updating databases, calling APIs, triggering functions not just chat or summarise.

The target is clear: solve friction for filmmakers, screenwriters, studio crews, or fans. Think of an agent that can manage script revisions, pull reference material, coordinate simple post-production steps, or help with production tracking. It has to run as a working system, not a demo that only works in a notebook.

Google for Developers highlighted the same idea in their recent post: agents that help creators write, produce, edit, and bring stories to life. The timing lines up with other experiments Google has run internally. One public write-up from the Google Cloud blog describes teams of agents that collaborated on short films, one agent handling the concept and style, another operating the generative tools, and a third handling editing and pacing. Those internal tests showed both the promise and the practical problems of getting agents to stay consistent across a full production pipeline.

To enter, you need four things: a hosted project URL, a public open-source repository with a clear license file, a three-minute demo video, and the completed Devpost form that marks which partner track you chose. The code must show the agent actually using Google Cloud services and the partner tool at runtime. Projects that only mention the tools in a README do not qualify.

Judging looks at four equal parts: how solid the technical implementation is, whether the design feels complete rather than a proof-of-concept, the potential real-world impact, and the quality of the idea itself. Teams can have up to four people. Most countries are eligible, though a short list of excluded territories appears in the full rules.

Resources are already up on the Devpost site. There are guides for Agent Builder (the lower-code path), the Python SDK for custom agents, and setup notes for each partner’s MCP server. Google Cloud free trial credits are available if you need them; the request deadline for the extra $100 credit is the end of August.

This is one of the more focused agent challenges so far. Most earlier ones stayed general. Here the problem space is locked to media workflows, and the requirement to take real actions through partner tools raises the bar past simple prompt engineering. Developers who already work with production tools or have built multi-agent systems will find the constraints useful rather than limiting.

Registration and full rules are live at the official Devpost page. The Google for Developers post that announced the open window is also easy to find if you want the short version of the pitch.