The rise of the agent runtime: The compute platform behind production agents

This came out while I was on holiday; some thoughts for The New Stack about how operationalising agents requires a different kind of infrastructure from what you need for familiar enterprise workloads.

You need an environment that looks a bit more like serverless with fast startup times, but also handles long-running, bursty, intermittent workloads that might scale to zero while waiting for a response and then need to restart just as fast.

You still need identity and security; that's more complicated because you need to both give agents access and potentially protect your systems from them because they run untrusted code relentlessly to achieve the goal you've given them.

You definitely want external gateways applying governance so agents can't hack the sandbox to achieve goals (see relentless!). Plus you need the agent to be able to take actions, in your existing business systems, using your existing workflows: so access to tools and services you want agents to use.

The new Azure Container Apps Sandboxes and Azure Container Apps Express are custom made for those needs, Microsoft tells me. The Foundry team already built the Microsoft Foundry Agent Service on top of them and organisations who want to make their own agent platform can do that too.