Two releases move AI agents closer to where the work happens
Mistral opened a hosted platform for building agents, and Microsoft's local toolkit can now run a coding model on your own machine.
Kivorent News desk, , 2 min read
What to know
- Mistral opened a public preview of managed agents in its Studio platform on November 5.
- Microsoft's Foundry Local can run Alibaba's Qwen3.8-27B coding model on a local machine and connect it to GitHub Copilot.
- One release puts agents on the vendor's servers and the other puts the model on yours. Which fits depends on your data and your hardware.
Two announcements last week point in opposite directions and are worth reading together. Both are about AI agents, the systems that take a goal and work through it step by step using tools. One runs them in the vendor's cloud. The other brings the model onto your own computer. Our account of both comes from AI Weekly's reporting on the companies' announcements.
Mistral: agents as a hosted service
Mistral, the French AI company, opened a public preview of managed agents in its Studio platform on November 5. The release has several parts:
- Managed agents. Multi-step agents that can use custom skills, run code in a sandbox and hand work to subagents. They run on infrastructure Mistral operates.
- An Agents SDK for building them in code.
- A visual workflow builder for people who would rather draw the steps.
- An apps command-line tool and evaluation tools for packaging agents and testing how well they do.
Private cloud and on-premises deployment are due in late November. That matters to banks and other regulated companies that cannot send data to a shared service. AI Weekly's report lists BNP Paribas, Accenture, Capgemini and Tata Consultancy Services among the companies already using the platform.
Microsoft: a coding model on your own machine
Microsoft's Foundry Local, its toolkit for running models on a PC, can now run Qwen3.8-27B, a coding model from Alibaba. It connects to GitHub Copilot through the option that lets developers bring their own model, so the familiar editor experience stays the same while the model runs locally.
AI Weekly cites published benchmark scores of 73.0 on Terminal Bench 2.1 and 61.7 on SWE-bench Pro, two tests of how well a model completes real programming tasks. Benchmarks are a starting point. How a model handles your own codebase is the test that counts.
How to choose between them
The trade-off is the usual one between hosted and local AI.
| Hosted agents | Local model | |
|---|---|---|
| Where your code and data go | To the vendor's servers | Nowhere |
| Hardware you need | A browser | A machine with enough memory to hold the model |
| Who maintains it | The vendor | You |
| Cost | Usage fees | Hardware and electricity |
A team with strict rules about source code leaving the building has a new reason to try the local route. A team that wants agents running long jobs without babysitting a workstation will find the hosted route easier.
We cover the trade-off in more depth in on-device AI versus cloud AI.
Sources
Spotted an error? Tell us and we will fix it.