Mark Cavage on Agentic AI, Sandboxing & Enterprise Security | Ep 10

Why It Matters

Agentic AI promises unprecedented automation, but without sandboxing and cultural discipline, enterprises face security and operational risks; Docker’s roadmap illustrates how platform leaders can capture value.

Key Takeaways

• Agentic AI needs sandbox mode to prevent harmful autonomous actions.
• AI agents can act as virtual product and engineering teams.
• Historical cloud building teaches leveraging past research for new platforms.
• Cultural alignment is critical when merging diverse engineering teams.
• Docker aims to create a new agentic platform built on Unix principles.

Summary

In episode 10 of Future of Data and AI, Docker COO Mark Cavage talks about the rise of agentic AI, the need for sandboxed “yellow mode” execution, and how enterprises can secure autonomous agents.

Cavage describes AI agents acting as virtual product managers and engineers, competing on scope and shipping speed, and how this can boost productivity but also create “addiction” to constant feature releases. He stresses that unrestricted agents can cause damage, so permission‑checking must be toggled.

He draws parallels to his early cloud‑building days at AWS and Oracle, noting that studying legacy systems—from 1960s mainframes to 1990s grid computing—provided reusable primitives. He also highlights cultural alignment as the toughest hurdle when integrating teams from different backgrounds.

The discussion signals a shift toward platforms built on Unix‑style micro‑agents, with Docker positioning itself to deliver the next generation infrastructure for the agentic era. Enterprises that adopt secure sandboxing early will gain a competitive edge while mitigating AI‑driven risk.

Original Description

He helped build the infrastructure that runs the modern internet. First AWS. Then Oracle Cloud Infrastructure. Then Heroku. Then Stripe. Now he's at Docker — and he thinks we're about to need a completely new layer underneath all of it.

When Mark Cavage, President & COO of Docker, joined the company, the question wasn't whether agents were coming. It was whether the infrastructure underneath them was ready.

It wasn't.

Before anyone was talking about agentic workloads in production…

Before AI tools started writing, running, and deploying their own code…

Before CISOs had a framework for reasoning about autonomous systems…

There was a simple but uncomfortable realization:

Containers were built for immutable, predictable software. Agents want to mutate everything.

In this episode of the Future of Data & AI Podcast, Mark Cavage — President & COO of Docker and one of the founding engineers of Oracle Cloud Infrastructure — joins Raja Iqbal for a candid conversation about what the agentic era actually demands from infrastructure.

Mark has spent over two decades building the systems that power modern cloud. Through Docker, he's now working on the sandbox layer that lets enterprises deploy agents at scale — without handing over control to a system nobody fully understands yet.

This conversation goes beyond the hype.

What You'll Discover:

🔹Why containers alone aren't enough for the agentic era.

Containers were built for immutable software. Agents mutate, write, and act — and Mark explains exactly what breaks, and what Docker built to fix it.

🔹What YOLO mode actually means — and why it matters.

Agents running without a human in the loop sounds reckless. Mark explains why that's actually the goal, and how the micro VM sandbox makes it safe enough for enterprise.

🔹The 1000x risk surface no one is talking about.

Every AI-generated pull request, every "authored by Claude" commit, every autonomously deployed dependency is stacking security debt. Mark breaks down what that means for your CISO.

🔹Trusted MCP servers and Docker Hardened Images.

What they are, why they exist, and why supply chain security for AI tools is about to become one of the most important conversations in enterprise tech.

🔹Mark's bets for the next 12 months.

CFOs demanding ROI on token spend, the open source project that no one is talking about, and what the future of Agentic AI looks like.

This episode is for:

🔹ML engineers and DevOps teams building with agents

🔹CISOs and security leaders managing a 1000x larger risk surface

🔹Platform and infrastructure leads evaluating MCP servers and supply chain security

🔹CTOs and engineering leaders figuring out what "agentic" actually means for their org

🔹 Founders deciding where the next infrastructure layer gets built

This isn't a conversation about demos or roadmaps.

It's about the infrastructure that agents actually need to run safely, reliably, and at scale; and whether the industry is building it fast enough.

If you're deploying agents in production, managing the security conversation, or trying to understand where Docker fits in the agentic stack… this episode is worth your time.

For more episodes: https://www.youtube.com/playlist?list=PL8eNk_zTBST_jMlmiokwBVfS_BqbAt0z2

For highlights: https://www.youtube.com/playlist?list=PL8eNk_zTBST-YYNgPcw3rO9Tvn7fEjR4A

Visit our podcast page for more info: https://datasciencedojo.com/podcast/