Build AI agents with the Mistral Agents API
Mistral AI
Mistral just launched an Agents API that lets its models run code, search the web, generate images, and remember past chats. It's Mistral's pitch for real enterprise agents, not just chatty text generation.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral AI wants you to stop thinking of its models as glorified autocomplete. The company's new Agents API, released today, bundles its language models with built-in connectors for code execution, web search, image generation, and MCP tools, plus persistent memory across conversations. The pitch is simple: raw language models are great at writing paragraphs but bad at remembering what happened five minutes ago or actually doing something in the real world. This API is Mistral's answer to that gap.
The numbers Mistral is showing off are hard to ignore. On the SimpleQA benchmark, Mistral Large jumps from 23% accuracy to 75% once web search is bolted on. Mistral Medium goes from 22.08% to 82.32%. That's not a marginal bump — that's the difference between a model guessing and a model actually checking its work against the internet before answering.
Mistral is also leaning hard into the agent-orchestration angle, where multiple specialized agents hand tasks off to each other inside a single conversation. Their cookbook examples show a Github-connected coding assistant supervising a DevStral-powered developer agent with full repo access, a financial analyst agent pulling data across several MCP servers, and a Linear-integration bot that turns raw call transcripts into product requirement docs and actual tickets. There's also a travel planner and a nutrition tracker, which feel more like demo-day flourishes than the real selling point here.
Underneath all of it sits a stateful conversation system, so developers don't have to manually track history or rebuild context every time — you can branch a conversation, resume it later, or stream responses in real time. Mistral is positioning this less as a chatbot upgrade and more as infrastructure: the backbone, in their words, for enterprise platforms that need AI to actually do multi-step work, not just describe it.
Whether enterprises bite depends on whether these connectors and handoffs hold up outside a curated cookbook demo, where MCP servers are already tidy and APIs behave. Mistral has built the scaffolding. Now it needs customers stress-testing it in messy, real production environments.
My take — AI-written commentary, not fact-checked reporting
I like that a European lab is shipping real agent infrastructure instead of just chasing benchmark headlines for the millionth chatbot release. The web search numbers alone — 23% jumping to 75% accuracy — say more about the current state of ungrounded LLMs than any safety paper I've read this year. My skepticism is reserved for the orchestration layer: handoffs between agents look clean in a cookbook, and they usually fall apart the moment real users start typing nonsense into them.
Read more about this at: Mistral AI