Mistral AI has launched a new Agents API that enables organizations to build their own AI agents to automate their business processes. With this API, the French AI startup is targeting enterprise customers who want to automate quickly without complex AI processes. The API combines the new Medium 3 model with tools for web search, code execution, and document processing.
Developments in AI agents are happening fast. AI agents are interesting, but many organizations struggle with where to start. Mistral AI aims to lower that barrier with an API that enables companies to deploy AI agents without requiring extensive programming knowledge.
With the Agents API, developers can relatively easily add autonomous AI capabilities to existing applications. It uses Mistral’s new Medium 3 model as a basis and offers various built-in connectors for common business tasks.
Web search improves accuracy
A notable improvement is in the web search functionality. In tests with the SimpleQA dataset, Mistral Large’s accuracy increased from 23% to 75% when web search was enabled. For Mistral Medium, the improvement was even greater: from 22% to 82%. These figures demonstrate that access to up-to-date information significantly enhances AI performance.
The API includes several tools that can be used directly for enterprise applications. Code execution allows Python scripts to be executed securely, while image generation uses Black Forest Lab’s FLUX1.1 model. For document processing, companies can integrate their own document libraries into Mistral Cloud.
Orchestration of multiple AI agents possible
A key feature is the ability for multiple agents to work together. Organizations can deploy specialized agents for different tasks and have them communicate via handoff mechanisms. Mistral supports the MCP protocol introduced by Anthropic for this purpose. This aligns with the trend toward multi-agent systems, which are becoming increasingly important. Google’s Agent2Agent protocol is not yet supported.
Mistral has developed various use cases, ranging from agents that integrate GitHub and assist with programming to travel agents and agents that provide nutritional advice. The API also supports stateful conversations, in which agents remember context over longer conversations.
Proprietary model without open source
One caveat is that Medium 3 is a proprietary model, unlike previous open source releases from Mistral. This means that organizations are dependent on Mistral’s platform for access. For the AI community, which values transparency, this shift raises questions.
The pricing structure reflects the enterprise focus. In addition to the standard model cost of $0.40 per million input tokens, there are additional costs for connectors: $30 per 1,000 calls for web search and code execution, $100 per 1,000 images for image generation. These prices can quickly add up with intensive use, but are quite competitive compared to other providers.
For organizations that view AI agents as inevitable, Mistral’s API presents an interesting option. It combines ease of use with enterprise-grade features.