Is your organization AI ready? Take the 3-minute test

All articles

Nov 19, 2026 · 5 min

How Kindred Voice Chose CreateOS Sandbox for Secure Code Execution in Enterprise Voice Agents

Kindred Voice CEO Phillip Hwang on why his team chose CreateOS Sandbox for secure code execution: an isolated microVM per run behind an egress allowlist.

NNaman Kabra

How Kindred Voice Chose CreateOS Sandbox for Secure Code Execution in Enterprise Voice Agents
On this page

Case study | CreateOS Sandbox

By Naman Kabra, co-founder, CreateOS

Kindred Voice chose CreateOS Sandbox for secure code execution because it provides an isolated microVM per run and a host-enforced network allowlist. This gives the team a runtime for future agent-written code without building its own sandboxing layer, so its engineers can stay focused on voice.

At a glance

  • Company: Kindred Voice
  • What they build: enterprise voice agents that answer inbound calls for large companies, built from each customer's own operating procedures
  • The need: secure code execution for AI agents, meaning running untrusted or agent-written code next to customer data
  • Why CreateOS Sandbox: an isolated microVM per run with its own kernel, behind a network egress allowlist the host enforces
  • What it unlocks: no in-house sandboxing layer to build, engineers who stay on voice, and a direct answer for enterprise security teams
Phillip Hwang, CEO of Kindred Voice, on why Kindred chose CreateOS Sandbox. Full interview, 77 seconds.

I filmed this story with Phillip Hwang, CEO of Kindred Voice, at our studio in October. His advice to other teams building agents was simple:

"If your AI agents will run code for enterprise customers, choose your agent runtime before your customers ask."

Phillip Hwang, CEO, Kindred Voice

About Kindred Voice

Kindred builds speech-to-speech voice agents for large companies. The agents answer inbound calls, and each one is built from the customer's own operating procedures. These are high-stakes conversations, so Kindred holds its agents to a strict bar: they have to follow each client's workflow exactly and prove they are reliable. In Phillip's words, the agents work inside customers' systems and follow each customer's process step by step.

That is what makes Kindred interesting to me. Their agents do not just talk. They act.

The challenge

Agents that act will need code execution

As Kindred's agents take on more of the work inside customer systems, the next step is code execution. Phillip put it plainly:

"As they take on more of that work, they will need code execution, which means running untrusted or agent-written code next to customer data."

Code an agent writes at runtime is code nobody reviewed. Running it next to sensitive customer data is the exact situation that needs a hard isolation boundary.

Enterprise security teams will ask

Kindred sells to large companies. Those buyers have security teams, and those teams will want to know two things about any agent that runs code: where does the code run, and what can it reach. A vague answer stalls a deal.

Building it in-house pulls engineers off the product

Kindred could build its own sandbox. That means owning isolation, networking rules, and the operational load that comes with them, all while the core product is voice. Kindred decided to solve this early. As Phillip said, "We needed secure code execution for AI agents before that happens."

Why Kindred chose CreateOS Sandbox

An isolated microVM per run

CreateOS Sandbox runs every execution in its own Firecracker microVM. Each run has its own kernel, so one run's code cannot share a kernel with another's.

"It gives us an isolated microVM per run, with its own kernel, so runs don't share a kernel with each other."

For untrusted and agent-written code, that per-run boundary is the whole point. One customer's run stays in one microVM.

A network egress allowlist the agent cannot change

Isolation is half the story. The other half is what the code can reach. Each run can sit behind a network egress allowlist that the host enforces, outside the run itself.

"Agent code reaches only the hosts we approve, and nothing inside the run can change that list."

Even if agent-written code misbehaves, it cannot widen its own network access.

Built for running untrusted agent code

The same runtime covers what agent teams tend to need next. Runs can be paused and forked, and teams can bring their own bucket, and on enterprise plans their own compute, so data stays where their customers expect it. Kindred's decision centered on the two guarantees above: per-run isolation and a host-enforced egress allowlist.

What it unlocks for Kindred

  • No sandboxing layer to build. Phillip told us Kindred won't have to build its own sandboxing layer.
  • Engineers stay on voice. The team keeps its focus on the product its customers buy.
  • A direct answer for security reviews. The isolation story fits in one sentence.

"When an enterprise security team asks where our agents will run code and what that code can reach, we'll have a direct answer: an isolated microVM per run, and only the hosts we allow."

Phillip Hwang, CEO, Kindred Voice

That line is the one I keep coming back to. Most teams find out they need an agent runtime when a customer's security questionnaire lands. Kindred made the call first.

What's next

Kindred's agents will keep taking on more work inside their customers' systems, and code execution will become part of that work. When it does, the runtime is already chosen. We are excited to keep working with Phillip and his team as that happens.

Phillip closed the video with this:

"For secure code execution for AI agents, we chose CreateOS Sandbox."

Try CreateOS Sandbox

If your agents will run untrusted or agent-written code for enterprise customers, give CreateOS Sandbox a try. Every run gets an isolated Firecracker microVM, pause and fork, and the option to bring your own bucket, and on enterprise plans your own compute. Each run can sit behind a network egress allowlist that the host enforces. Start a sandbox and see how it fits your agent stack.

Keep reading

All articles

Bring us one stuck workflow.

We take it to governed production with you, on your systems, in weeks, not quarters.