AI agents

AI agents represent a modern application building approach based around artificial intelligence and neural networks. Agents expand the capabilities of large language models and lay the groundwork for building chatbots, assistants, and voice interfaces, thus helping to automate various routine tasks.

Agents comprise four key components:

• LLM: Actual language model with fixed setting.
• Prompt: Describes the agent's behavior and role.
• Tools: Allow the agent to leverage external capabilities, such as APIs, functions, or internet search.
• Memory: Stores context and interaction history.

Such architecture enables the creation of text and voice agents that behave more naturally and autonomously than classic chatbots.

Developing agents in AI StudioDeveloping agents in AI Studio

AI Studio has everything you need to create AI agents: models that support function calling and response formatting, ready-made customizable RAG and internet search tools, as well as MCP Hub that allows connecting external APIs via MCP servers.

AI Studio allows you to create agents in various ways:

• In AI Playground.
• Using the Yandex Workflows specification constructor.
• Using the Responses API or Realtime API.
• Via open source frameworks, for example, OpenAI SDK, LangGraph, or LangChain.

ToolsTools

Agents can automatically invoke tools to get additional info for generation or perform the necessary actions. AI Studio comes with these bundled tools:

• Retrieval Tool implements the RAG scenario and allows the AI agent to search through your files (knowledge base) for the information for its response. You can upload your knowledge base documents in the management console, via the Vector Store API, or via the Files API and create a search index. Search indexes store information from your documents in vector form and allow agents to use it to respond.
• WebSearch Tool allows the agent to search for information on the internet via the Yandex search base to enrich its responses with up-to-date query-related information.
• MCP Connector is responsible for connections to MCP servers to work with third-party APIs.
  In the MCP Hub section, you can create and set up connections to new and existing MCP servers and monitor their status.

Agent creation APIAgent creation API

AI Studio provides two OpenAI-compatible APIs for development of different types of agents. Both APIs save data on client states between requests and basically perform the same functions of connecting models, tools, and memory, but are optimized for different types of interactions depending on the agent.

• The Responses API is an API for text scenarios. Agents created using the Responses API take note of the conversation context and can automatically invoke connected tools. The Responses API supports any models.
• The Realtime API is an API for voice scenarios. This API is designed works with specialized multimodal models accepting audio input and synthesizing an audio response. The Realtime API supports all available tools, including Retrieval and WebSearch.

Workflow constructorWorkflow constructor

The workflow constructor allows you to build complex AI-based scenarios using out-of-the-box automation and management steps. The constructor will suit those of you who prefer visual editors and low-code platforms.

You can learn more about workflows and automation steps in the Yandex Workflows guide.