[AI in Action] Gemini Agentic Video: 4 Hidden Prerequisites and My LINE Bot Integration Process

In late August, Google introduced agentic video in Gemini, a feature that allows users to ask questions about videos instead of just receiving summaries. However, there are four hidden prerequisites that must be met for agentic video to work on Vertex AI:

1. api_version must be set to v1beta1. If left at the default v1, agentic mode will be disabled.

2. media_processing must be set to AGENTIC. If not specified, the system defaults to STATIC mode.

3. The model used must be one of the supported options: gemini-3.7-flash, gemini-3.6-flash, or gemini-3.5-flash-lite. Using an unsupported model will silently revert to STATIC mode.

4. The thinking_level parameter must be set. This is the most critical prerequisite, as failing to set it correctly will result in the API returning a 200 response, but the agentic mode will not be activated, leading to unexpected costs.

In addition to these four technical prerequisites, there are also two environmental prerequisites to consider:

1. The SDK version must be 2.20.0 or higher, as the media_processing field was added in this version.

2. The documentation examples provided by Google are written for the Gemini Developer API, not Vertex AI. The behavior on Vertex AI is different, requiring a stateless re-query approach instead of the multi-turn conversation mechanism used in the Developer API.

These hidden prerequisites can be easily overlooked, leading to unexpected costs and performance issues. The author of the article created four tests to ensure each prerequisite was met, as failing to meet any of these requirements would result in the program running correctly, but with incorrect costs and performance.