Google Cloud’s Gemini Enterprise Agent Platform documentation, last updated Sept. 15, 2026, describes how customers can run reinforcement-learning fine-tuning on Gemini models from the console and the API. The pages sit under Models > Tuning. They are product docs, not a music-industry announcement. They matter to any rights, catalog or touring tool that already lets software agents act under a Google Cloud identity.
The overview page says reinforcement learning “lets you fine-tune Gemini by using your own prompts and self-defined reward functions.” It says the method can “boost the reasoning capability and output quality of Gemini models for your own use cases and agentic workflows.” Supported hardware for that pass, the same page lists, is Gemini 3.5 Flash. Tuning endpoints named there are us-central1 and europe-west4. Tuned models serve from the matching multi-region endpoint. The API path in the examples is v1beta1.
A job has three parts. A JSONL dataset in Cloud Storage. Hyperparameters. One or more reward functions that score each reply. Every dataset line carries a system instruction, prompt contents and a references map the scorer can read — a ground-truth string, a rubric, a test fixture. Contents may include text or file data. File types listed include image, audio and video, inside the service’s dataset limits. Audio in the schema is the only music-shaped hook in the docs. It is a training example format. It is not a license to a recording.
Rewards clip to the range -1 to 1. Types listed: string matching, code execution, a Gemini autorater, a Cloud Run service the customer owns, or a composite of those with weights, up to 16 scorers. Cloud Run scoring needs the Secure Fine Tuning Service Agent — service-PROJECT_NUMBER@gcp-sa-vertex-tune.iam.gserviceaccount.com — to hold invoke rights on that service. If a music-tech shop scores “did this draft respect the split sheet,” that score is their function. Google’s page does not define neighboring rights.
Console path: Create tuned model, pick Reinforcement learning fine-tuning, set rewards, point at the bucket, set epochs, learning-rate multiplier, samples per prompt, LoRA adapter size, batch size, thinking level MINIMAL or HIGH. Continuous tuning is documented the same day. Patterns named: supervised fine-tuning into RL, or RL into more RL. Start from a prior tuned model or a checkpoint. Use listed: more data after underfit, refresh on new examples, further customize. Job states include pending, running, succeeded, failed, cancelling, cancelled. Charts show reward, generation length, reward latency.
That is the model shop. The identity shop is separate. Google Cloud’s Policy Troubleshooter documentation tells operators to give agents their own principal, not a shared user login. The remote MCP server for Policy Troubleshooter, the company says, “provides your agents direct access to tools that can help troubleshoot Identity and Access Management issues and errors.” It also says: “We recommend that you create a separate identity for agents that are using MCP tools so that access to resources can be controlled and monitored.” API keys are not accepted. Example prompts on that page ask why a named principal was denied storage.objects.get, or what a permission-error ID means. For a catalog service, a denied read on a stem bucket or a metadata table is no longer a ticket in the IT queue. It is the product failing in front of a manager.
None of those pages mention labels, publishers, SoundExchange or a DSP. The operational read for a music company on Google Cloud is narrower. If an in-house agent drafts metadata, files a cue sheet, or opens a ticket against a master, and that agent authenticates as its own IAM principal, then RL tuning changes how the model is steered and Policy Troubleshooter changes how a deny is explained. Reward design is the rights risk. A scorer that pays the model for speed will not pay it for holding an unreleased title. That trade-off is not in Google’s docs. It belongs in the customer’s reward function.https://docs.cloud.google.com/release-notes
https://www.theepochtimes.com/tech/ai-agents-cheated-in-google-experiment-researchers-report-6087294