AI agents need cheaper sandboxes | Alex Jung from Unikraft (Heavybit DevGuild)
Unikraft can schedule an application in roughly 10 milliseconds and move an idle workload from memory to cheaper NVMe storage—economics that become increasingly important as agents spawn more agents and consume more sandboxes.
Alex Jung is the co-founder and CPO of Unikraft, which grew from an academic unikernel project into a cloud platform for fast, resource-efficient workloads.
Recorded at Heavybit DevGuild, Alex explains why Unikraft's original pitch failed even though its open-source community attracted an investor, why deploying lightweight VMs on conventional cloud infrastructure did not deliver the intended economics, and how agent sandboxes are creating new demand for fast startup, scale-to-zero and dense workload consolidation.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
What is a unikernel?
A unikernel is a highly specialized virtual machine. A general-purpose system such as Linux supports many workloads, while a unikernel packages an application with only the operating-system components it needs. The result can retain VM isolation while improving performance and using fewer resources because the unused parts of a full system are absent.
Why was Unikraft difficult to pitch at the beginning?
We were explaining a highly technical compute model to non-technical audiences without a clear product or market wedge. We could promise fast, secure and cheap workloads, but 'all of compute' was not a credible market plan, and our benchmarks came from research environments rather than noisy production systems. After enough pitch events, we started asking ourselves what exactly we were selling.
What convinced Unikraft’s first investor?
The pitch itself was poor, but the project happened to trend on GitHub and we had invested in an open-source community. A Discord server, careful README and website, university workshops and small hackathons showed technical momentum. An infrastructure-focused pre-seed investor found us and backed the founders and community even though the business graph in the deck was rudimentary.
Why are AI agents a natural use case for Unikraft?
Agents need isolated environments in which to execute and build, and a sandbox can be exposed through an HTTP API rather than only through SSH. The number of agents and harnesses is growing, with agents increasingly spawning other agents. Software generation is expanding faster than the available hardware, so reducing the resources and startup time required for each environment becomes valuable.
Why couldn’t you simply run the lightweight VMs on AWS?
We had little visibility or control over how quickly EC2 would schedule a virtual machine. More importantly, a specialized application might need only about 16 MB of memory while the smallest available instance allocated roughly 512 MB. The application was efficient, but the surrounding cloud product forced us to pay for and reserve an order of magnitude more capacity than it needed.
What did Unikraft build instead?
During the pre-seed stage, we decided to build a cloud platform and control plane designed around micro-VMs from day one. It uses techniques including preloading, read-only memory and snapshots to schedule an arbitrary application in roughly 10 milliseconds. Owning that layer lets the product preserve the properties that made the unikernel useful in the first place.
How does scale-to-zero work?
The platform can observe CPU and network activity. When an application is no longer doing useful work, it takes a snapshot and moves the state from expensive memory to cheaper NVMe storage. Because startup is measured in milliseconds, a request can wake the workload again while the platform uses its active memory and CPU for something else in the meantime.
Which customers benefit most from that model?
It is especially useful for products running many unevenly active workloads. A DevTool's free-tier users may provision something once and rarely return, while a smaller share of customers produces most of the traffic. Those idle and lightly used environments still need to respond when called, but they should not permanently occupy the same resources as active, paying workloads. Dense consolidation improves that economics.
How does Unikraft make money?
There is a hosted platform for individual users as well as larger deployments for companies operating at scale. The larger customers gain the most from consolidating many low-utilization services onto fewer machines. Pricing is roughly based on the number of servers and virtual CPUs used.