What the Infrastructure Layer Covers
3 entriesAn agent is more than a model call. It runs for minutes or hours, executes code it wrote, drives browsers, calls tools with a user's credentials, remembers earlier sessions and must survive failures. Managed providers sell each of these pieces as a service so teams do not have to build and secure them. This page groups providers by layer. Many span several layers, so check each provider's documentation for what is included.
Hyperscaler Agent Runtimes
12 entriesManaged services from the large clouds for hosting, scaling and governing agents.
Sandboxes & Code Execution
12 entriesIsolated environments where an agent can run code and commands safely.
Serverless & GPU Compute
8 entriesOn-demand compute for model hosting and agent workloads.
Cloud Browsers & Computer Use
7 entriesManaged browsers that agents control to read and act on the web.
Tools, Identity & Gateways
8 entriesManaged access to third-party tools with the right credentials.
Memory & State Services
7 entriesManaged memory so agents keep context across sessions.
Durable Execution & Orchestration
7 entriesEngines that keep long-running agent workflows alive through failures.
Related Pages
7 entriesNeighbouring pages on this site.
Frequently Asked Questions
What is the managed infrastructure layer for AI agents?
The set of cloud services that sit under an agent: a runtime to host it, sandboxes for code it writes, compute for models, browsers it controls, tool and identity services, memory, and durable execution. Providers offer these separately or as a bundle.
Do I need a sandbox if I use a hyperscaler agent runtime?
Often yes. A runtime hosts the agent itself, while a sandbox isolates code the model generates. Some runtimes include a code interpreter, so check what each one provides.
How should I choose between providers?
Match them to your existing cloud, data location and compliance needs, then compare startup time, isolation model, session length limits, pricing model and observability. Test with your own workload.
Can I mix providers?
Yes. Teams commonly host an agent on one cloud, run code in a specialist sandbox and use a separate memory or tool service. Keep interfaces such as MCP and open protocols in mind to limit lock-in.
Are there security considerations?
Yes. Agents run model-generated code and hold credentials, so use isolated sandboxes, short-lived scoped credentials, network egress controls and audit logs.
How current is this page?
Links were checked in October 2026. This market is young and product names, pricing and availability change quickly, so confirm details on each provider's site. Sites that block automated checks are kept and may need a manual click. Descriptions summarize each provider's own documentation and are not rankings.