
Akamai shipped two beta products on October 27: a managed inference service and an agent sandbox. The inference side is predictable — twelve open-weight models, OpenAI-compatible APIs, serverless billing. The sandbox is where it gets interesting. It uses WebAssembly to isolate code that agents write and execute, keeping generated code sequestered from the agent runtime itself. After a week of headlines about agents escaping microVMs and AI tools leaking agent logs, the timing is pointed.
The Sandbox Architecture
The design is a two-tier isolation model. Your agent application runs inside a Spin WebAssembly environment on Akamai Functions. Code the agent generates — the JavaScript or Python it writes and executes — runs in a more restricted nested WASM environment inside that application. Two layers, with the inner one deliberately tighter.
Outbound connections from generated code are blocked by default. If the agent needs to call an external API, that must be explicitly permitted. State between execution steps is preserved via key-value storage. Each step runs in roughly one-minute intervals during the preview period.
The architectural intent is sound: separating the agent’s runtime from the code it produces means that if generated code is malicious or buggy, it cannot directly reach the agent’s credentials, upstream state, or your infrastructure. This is the conceptual split most agent sandboxes miss.
WebAssembly vs. MicroVMs: The Honest Trade-off
Akamai chose WebAssembly over Firecracker or gVisor. That is a deliberate bet with real trade-offs. WebAssembly cold starts are in the microseconds. Firecracker — what E2B and OpenAI’s sandbox use — starts in roughly 125ms and gives each workload its own kernel. That kernel boundary is what makes microVMs the current gold standard for truly untrusted code: a kernel exploit inside one VM cannot escape to the host or adjacent workloads.
WebAssembly does not provide that. It offers process-level isolation without a separate kernel. For most coding agents, test runners, and automation workflows, that is sufficient. For running completely untrusted code from unknown sources — think multi-tenant platforms where users submit arbitrary scripts — microVMs remain the stronger boundary.
The upside Akamai is betting on: WebAssembly runs everywhere across their 4,350+ edge locations with microsecond cold starts and a tiny resource footprint. That is a real advantage for latency-sensitive agents operating globally. The Spin SDK powering this is open source — Akamai acquired Fermyon, its creator — and the framework already underlies similar architectures at Cloudflare Workers and Fastly Compute.
What Is in the Inference Catalogue
Akamai Inference launches with twelve open-weight models covering 8 billion to 2.8 trillion parameters. Named models at launch: NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b, Google Gemma 4, and embedding models for retrieval-augmented generation. The API is OpenAI-compatible, meaning existing integrations migrate without rewrites.
One honest note from the docs: tool calling and structured outputs “need individual testing.” That caveat is worth heeding before routing production agent workflows through here. Hardware is serious — Akamai is running RTX PRO 6000 Blackwell Server Edition and Blackwell Ultra clusters with 2.3 TB of HBM3e memory per node. NVIDIA Vera Rubin infrastructure is planned for 2027.
What Is Still Missing
Beta means beta. There is no complete pricing sheet — inference is free during the beta period, but dedicated deployments will charge by GPU-hour including idle time, and serverless token rates are not yet published. The compatibility matrix for specific API features is incomplete. Additional access controls for agent permissions are listed as planned but not yet available.
This is a real product worth watching, not a press release without substance. But production agent workloads should wait for at least one or two more release cycles. Anthropic’s sandbox-runtime, also in beta, is worth evaluating in parallel if you need agent isolation today.
How to Get Access
The Spin SDK is open source at github.com/fermyon/spin — you can evaluate the WebAssembly isolation model locally before committing to Akamai’s hosted version. For Inference beta access and Blackwell cluster access, contact Akamai directly. The official announcement has the contact details and full model catalogue.

