Agent Studio provides a collaborative workspace for discovering models, refining system instructions, and optimizing prompts. With built-in tools for natural language refinement and side-by-side comparison, you can efficiently prototype and test generative AI solutions.
This page outlines the key features implemented within Agent Studio that facilitate transforming innovative ideas into generative AI solutions, for example, solutions that benefit customer service and support, content creation, healthcare, financial services, and software development.
Roles
You can control which Google Cloud services that users can access through Agent Studio by assigning the following roles:
| Role | Description |
|---|---|
| Administrator | Responsible for Google Cloud setup that interacts with agents using technical language to connect Agent Platform with the Google Cloud ecosystem. |
| Builder | Includes the ML developers and App developers, who use the developer tools. Agents use developer-specific language with a scope limited to Agent Platform. |
Features and experiences
This section describes Agent Studio capabilities.
Navigation and playgrounds
Agent Studio opens on the prompt workspace, where you write prompts, adjust model settings, and compare responses. The left navigation gives you the following destinations:
| Destination | Description |
|---|---|
| New | Opens a new playground. See the following table for the available playgrounds. |
| Agents (preview) | The list of agents you have designed. From here you can open an existing agent or create one. For more information, see Design agents in Agent Studio. |
| App builder (preview) | A workspace for generating a web application from a description, and for editing and deploying the generated code. |
| Gallery | Sample prompts, apps, and agents that you can open and run. Filter the samples by task, media type, or feature. |
| Documentation | Opens the Agent Platform documentation in a new tab. |
To start fresh work, click New in the left navigation and choose a playground. Each playground is tuned for one kind of model output:
| Playground | Description |
|---|---|
| Chat | Multi-turn text and multimodal prompting. Supports the full set of model settings, grounding options, and built-in tools. |
| Image | Image generation and editing. For more information, see Generate images. |
| Video | Video generation. For more information, see Video generation overview. |
| Music | Music generation. For more information, see Music generation overview. |
| Speech | Speech generation and transcription. For more information, see Text-to-speech and Speech-to-text. |
| Live API | Low-latency, bidirectional voice and video interaction. For more information, see Live API. |
The Image, Video, Music, and Speech playgrounds are collectively known as Agent Media Studio.
Onboarding and administrative
The onboarding and administrative features let you access and set up your environment.
| Feature | Description |
|---|---|
| Get API key | You can access your API key quickly. |
| Getting started menu | A menu for new users with steps to properly activate their accounts, which include trying a prompt, adding a collaborator, and getting an API key. |
| Settings menu | A settings menu that includes options to switch between the previous and latest UI, change the appearance to use the light or dark system, generate and access API keys, and manage billing. |
Studio discovery and development
The studio discovery and development features enhance the developer workflow.
| Feature | Description |
|---|---|
| Interactive canvas view | A canvas that illustrates the artifact to be retrieved for development, which includes a setting/preview pane and a code view. |
| Minimap | A summarized history of prompt and response headers that can be turned off. |
| Slash commands | Type / in the prompt field to open the command menu. The
commands are grouped as follows.Prompt development:
|
| Side-by-side comparison | An updated comparison feature for models, system instructions, and prompts, which lets you compare multiple prompts and your responses simultaneously. |
| Upload large files | The ability to upload local or Cloud Storage files up to 50 MB. |
| Generate code files | A feature that generates application code files, which are then accessible in the canvas pane. |
| Download code files | The ability to download generated code files as a zip file. |
| Undo or Redo | Undo or redo functionality is available for actions like model settings and prompt changes. |
| Gallery | A gallery of sample prompts, apps, and agents. Filter the samples by task, media type, or feature. |
Prompt settings
The Model settings panel next to the prompt controls how the model generates a response. The available settings depend on the model and the playground you are using.
| Setting | Description |
|---|---|
| Thinking level | How much reasoning the model does before responding. For more information, see Thinking. |
| Output format | The response MIME type, which lets you request structured output such as JSON. For more information, see Control generated output. |
| Grounding with Google | Ground responses in Google Search results. Click Customize to configure the search behavior. For more information, see Grounding with Google Search. |
| Grounding with partners | Ground responses in a partner data source, such as Elasticsearch. |
| Grounding with your data | Ground responses in your own enterprise data using Agent Search or RAG Engine. |
| Function calling | Let the model call functions that you declare. For more information, see Function calling. |
| Code execution | Let the model generate and run code to answer a prompt. |
| URL context | Let the model fetch and analyze URLs that appear in your prompt. |
| Safety filter settings | Thresholds that control which responses are blocked. For more information, see Configure safety filters. |
| Output token limit | The maximum number of tokens the model can generate in a response. |
| Stop sequences | Strings that stop generation when the model produces them. |
| Seed | A fixed seed value, which makes responses more reproducible across requests that use the same prompt and settings. |
| Stream model responses | Return the response incrementally as it is generated, instead of waiting for the complete response. |
| Region | The region that serves the request. The available regions depend on the model you select. |
Built-in tools and capabilities
This section lists the built-in tools and capabilities within Agent Studio that automate a significant portion of the work for developers. Agent Studio integrates with Google Cloud, which provides an effortless experience that's collaborative and agentic.
| Feature | Description |
|---|---|
| Natural language refinement | For prompt refinement, you can describe how you want to change your prompt using natural language, and the system automatically optimizes the prompt for you. |
| Optimize system instruction | You can use a one-click feature that asks Gemini to automatically optimize a system instruction based on the prompt. |
| Refine prompt response | You can use the /prompt command to provide feedback and
to create a well-formatted revised prompt. |
| Help-me-write tool | You can use the /prompt command and a described intent to
create a well-formatted prompt, system instruction, and to recommend a
model. |
| Convert to agent (preview) | Turn the prompt you are working on into an agent. In the
Build with code menu, click Convert to agent. The model,
system instructions, prompt, and prompt variables carry over to the new
agent, which you can then design and deploy. For more information, see
Design agents in Agent Studio. The prompt or the system instructions must be non-empty before you can convert. This feature isn't available to signed-out users or to users without a billing-enabled project. |
What's next
Agent Studio quickstart: Send text prompts to Gemini
Learn how to send text prompts to Gemini using Agent Studio.
Deploy your prompt as a web application
Follow the guide to deploy your prompt as a web application from Agent Studio.