Agents API: Mistral publishes framework for the development of AI agents

Mistral AI has introduced Agents API, a framework that enables companies to create their own AI agents to automate tasks and processes. Mistral combines its own and external language models with a control system for specialized AI agents and a persistent memory to maintain the context between the various agents. The framework also has various connectors, for example for executing code or accessing documents. Agents API uses Anthropic's open Model Context Protocol (MCP) to connect to external applications.

AI agents can execute Python code in a sandbox

The framework currently includes four built-in connectors that AI agents can access at any time. In addition, users can also use them outside of agents directly via the chatbot. One such connector is the code interpreter, which can execute Python code in an isolated sandbox environment and check it for syntax errors. It is also suitable for mathematical operations as well as data analysis and visualization. Such a coding assistant can be created via a connection to GitHub and the Devstral language model, of which a new version was recently released.

Agents API also has a connector for image generation. Mistral uses the Flux 1.1 Pro Ultra language model from Black Forest Lab for this purpose. According to the company, this can be used to create visualizations for training material or graphics for marketing purposes. AI agents can access documents in the Mistral cloud via the document library. In this way, the connector drives the Retrieval Augmented Generation (RAG) by using the content of the documents as context for the output.

Web search supplements training data

Another knowledge base for the AI agents is the built-in web search, which supplements the training data of the language models with up-to-date information. On the one hand, AI agents can search the web independently based on the query; on the other hand, specific URLs can also be transferred to the connector to use the information on the page. In the SimpleQA benchmark, the quality of the answers provided by the web search in the Mistral Medium and Large language models increased threefold. Mistral Medium achieves a value of 75 percent with web search, while the large model achieves 82 percent.

To create their own workflow with several AI agents, developers first create all the required agents, each of which can access different language models and applications. There is no maximum number of usable agents in Agents API. Developers can then decide whether an AI agent delivers an output or forwards a request to another agent. This means that several AI agents can be linked together so that, for example, an agent for financial analysis outsources the research of stock market data to an agent for web searches and the calculations to another agent.

Detailed information on the Agents API can be found in the documentation. Mistral provides ready-made example agents on GitHub. Initially, the current versions of the Mistral Medium and Large language models are available. Support for other models is to follow. Most recently, JetBrains also published the Koog framework, an open-source software for the modular creation of AI agents.

(sfe)