Akamai launched two AI services in beta on 27 October: Inference, for running open-weight and custom models, and Agent Sandbox, for executing agent software and the code it produces. Together they give developers managed model access and a place to run the surrounding application on the same cloud provider.[1]
Agent Sandbox runs the agent itself in a WebAssembly component on Akamai Functions, using the open-source Spin framework. Generated JavaScript or Python runs in a separate, more restricted environment inside that component.[2]
Spin packages applications as WebAssembly components that can run locally, on self-hosted servers, in Kubernetes or through cloud services. That portability is useful during development: the framework is intended to preserve the application’s execution model as it moves between those settings.[3]
Akamai says the same Spin runtime can be used on a developer’s machine and on Functions. Key-value storage preserves context between steps.[2]
Connections are part of the application
In Spin’s HTTP interface, a component asks the host runtime to make an outbound request. Its manifest names the permitted destinations through an allowed-hosts setting. An attempt to contact another host fails. The configuration therefore becomes a concrete part of deploying an agent, alongside its code.[4]
For example, a component that reads an inventory API and calls a language model would need routes to those services. Moving it to a hosted environment means preserving both the executable component and the connections that make its task possible.[4]
The present preview has a specific rhythm: Akamai describes intermittent execution, with each step lasting about a minute before stopping until the next instruction. Longer-lived stateful components, able to run for days or weeks, are planned. Requests are routed to available instances in Akamai’s core cloud regions.[2]
On the model side, the Inference beta offers serverless access through an OpenAI-compatible API, plus embeddings for retrieval applications. The launch catalogue contains 12 open-weight models. Developers can therefore choose the model they call while using Spin to manage the application around it.[1][2]
Sources & further reading
- Akamai Launches AI Inference Service and Agent Execution Environment Akamai · 27 Oct 2026 · Accessed 28 Oct 2026
- Akamai Agent Sandbox Akamai · Accessed 28 Oct 2026
- Introducing Spin Spin project · Accessed 28 Oct 2026
- Making HTTP Requests Spin project · Accessed 28 Oct 2026