Summary

  • Akamai launched Agent Sandbox in beta on 27 October. Agents run on Akamai Functions; code they generate runs in a tighter JavaScript or Python boundary inside a WebAssembly sandbox, with outbound access blocked by default.
  • Akamai's product FAQ describes intermittent agents whose individual steps complete in about a minute and then stop. Components for agents that remain alive and retain state for days or weeks are a future addition.

A shift worker in a glass box

The new product has a sharp use case: let an AI agent write a little code, run it behind a wall, return a result and go quiet. Akamai's launch announcement places the agent on Akamai Functions and the code it writes in a second, restricted runtime. The product description says the generated JavaScript or Python starts with no outbound network access. The operator can grant selected connections and impose resource limits. That is a saleable control surface for a customer whose agents handle data belonging to many clients.

The time boundary deserves equal billing. Akamai says the technical preview supports intermittent agents: a step completes in about a minute, the agent stops, and another instruction starts the next step. It says stateful components for agents that stay alive for days or weeks will come later. The immediate product can help with bounded analysis, calculations and code execution. A standing worker that remembers a case, waits for an external event and resumes without someone stitching the context back together belongs to a different operating design.

Akamai's Functions documentation lists a default HTTP request-handler duration of 30 seconds, 128 MiB of memory and a 10 MiB request/response limit. The Sandbox product page describes a step of about a minute. Those published figures address different layers. A buyer needs Akamai's actual beta quota for the agent path and a test of how a long job is divided, saved and restarted. The access page offers a request form, while the Functions onboarding guide still describes a preview requiring approval.

The market bet is that control and placement make a short job valuable enough to buy. Akamai offers a global Functions fabric, a nested boundary for generated code and model access through compatible APIs. The customer supplies the workflow that ties separate steps into a service. For a SaaS team, that workflow includes tenant identity, permitted tools, state storage, retries and a way to prove which step produced which result. The sandbox can contain an execution; the customer still has to operate the job.

No public Agent Sandbox price appears on the cited product page. A published beta quota, rate card and tested state-handoff behaviour would turn the pitch into a procurement calculation. Until then, Akamai has a concrete product for short controlled work and a much longer sales conversation for the always-on agent.

Sources