Akamai’s agent sandbox: can AI run code safely now?

What happened

Akamai introduced two beta cloud services designed to operationalize enterprise AI: Akamai Inference and Akamai Agent Sandbox. Akamai Inference allows companies to deploy custom-tuned or open-weight models to process user requests, contract summaries, ticket classifications, and policy searches without managing foundational training environments.

The second release, Akamai Agent Sandbox, addresses the operational risks of autonomous software agents that execute code, call APIs, and trigger workflows. To isolate this activity, the agent executes via Akamai functions, while any code it generates runs inside a separate, locked-down environment. This mechanism creates guardrails to prevent agents from executing destructive commands across live corporate infrastructure.

Alongside the software offerings, Akamai announced an expansion of its compute capacity through Nvidia Blackwell Ultra GPU clusters. Building on its legacy of distributed web performance, Akamai aims to deliver low-latency inference by processing AI workloads closer to end users rather than relying strictly on centralized hyperscale cloud providers.

Why it matters

The enterprise AI landscape is shifting focus from raw model intelligence to operational execution. Deploying an agent that interacts directly with corporate tooling introduces severe operational exposure; an unchecked system operating with confidence can misfile records, misdirect sensitive communications, or execute unauthorized system commands. Isolating code execution inside an edge sandbox establishes a predictable boundary that limits an agent's blast radius when failures happen.

Akamai’s deployment of Nvidia Blackwell Ultra hardware at the network edge challenges traditional centralized cloud vendors. Running inference closer to users addresses enterprise latency demands and data control requirements. For organizations adopting operational AI, infrastructure that couples edge execution with physical isolation lowers the barrier to deploying active, task-oriented agents.

As execution guardrails mature, automated workflows will expand across back-office operations, administrative scheduling, claims, and customer support. Instead of basic informational chatbots, systems will gain the capability to execute refunds, reset credentials, and modify accounts. For operators, human value shifts decisively away from repetitive task execution toward workflow architecture, governance, and systematic exception handling.

What to watch next

• Whether enterprise adoption of isolated execution sandboxes accelerates the deployment of autonomous customer support actions, such as direct refunds and account modifications.
• How effectively Akamai's distributed Nvidia Blackwell Ultra clusters compete against centralized hyperscale cloud providers on latency and pricing for custom model inference.
• The emergence of standardized audit logging and permissioning frameworks across competing agent-execution platforms.
• How rapidly enterprise IT departments establish formal approval gates for agent-executed code versus traditional human review processes.

What this means for you

Review any existing or planned AI agent implementations against four non-negotiable operational criteria. Determine exactly where the agent executes its code, map the precise data folders and systems it can access, establish which actions require manual approval, and verify that comprehensive audit logs track every API call, record touch, and message. Never deploy autonomous tools under an unverified assumption that broad access equals productivity.

Implement a "draft and propose" workflow rather than allowing agents to execute external or critical operations autonomously. To stress-test internal resilience, ask engineering teams and vendors a single diagnostic question: if an agent makes a mistake at 2:00 AM, what is the worst action it could take, and what technical barrier prevents that failure? Base operational readiness on technical isolation and strict oversight rather than model confidence.

Key takeaways

• Akamai introduced beta services for Akamai Inference and Akamai Agent Sandbox, alongside an expansion into Nvidia Blackwell Ultra GPU clusters.
• Akamai Agent Sandbox isolates agent-generated code inside separate execution environments to prevent unauthorized or destructive operations across production systems.
• Akamai Inference enables enterprises to run open-weight and customized models closer to end users rather than relying solely on centralized cloud facilities.
• The operational focus of enterprise AI is shifting from model intelligence to managing the blast radius of autonomous agent actions.
• Effective AI governance requires shifting operational workflows to a draft-and-propose model supported by immutable audit logs and explicit permission boundaries.

FAQ

What is Akamai Agent Sandbox?

Akamai Agent Sandbox is an isolated execution environment designed to let autonomous AI agents run code safely. While an agent operates on Akamai functions, the code it writes executes inside a separate, locked-down boundary. This structural containment prevents an agent from damaging production environments, altering sensitive databases, or triggering unauthorized operations when attempting complex software tasks.

What is the purpose of Akamai Inference?

Akamai Inference is a beta service that allows organizations to run custom or open-weight artificial intelligence models on live user requests. Instead of supporting model training or foundational research, Akamai Inference handles operational tasks like summarizing contracts, classifying support tickets, drafting messages, and searching policy documentation directly on Akamai’s distributed cloud infrastructure.

What hardware is Akamai deploying to support AI inference?

Akamai is expanding its infrastructure by deploying dedicated Nvidia Blackwell Ultra GPU clusters. These high-performance graphics processing units provide the physical compute capacity required to run intensive AI workloads and inference tasks at the edge, offering enterprises faster processing speeds and localized data control close to end users.

What four questions should teams ask before deploying an AI agent?

Teams deploying an AI agent should ask where the agent runs its code, what exact data and systems it can access, which specific actions require human approval, and whether a complete audit log is maintained. Verifying these technical boundaries ensures that autonomous software operates within strict guardrails rather than exposing corporate systems to unmonitored risk.