Andreessen Horowitz is leading a $20 million seed round for Runta, a startup rebuilding the compute layer where AI agents actually execute. The round values Runta at more than $100 million, according to The Information and TNW.
Runta’s thesis is blunt: the existing cloud stack was designed for human developers, not autonomous software. Founder Guanlan Dai, who served as a technical lead on Cloudflare’s early edge proxy and caching team before building the core proxy at API startup Kong, argues that agents require a fundamentally different execution environment.
The “Agents Need a Computer” Argument
In a16z’s investment post, partner Martin Casado laid out the case. “AI agents just want a computer,” he wrote, meaning a full operating system install, fully stateful, capable of running locally or in the cloud, with security controls built in. He specifically distinguished Runta from the wave of sandbox cloud startups that address only a portion of the isolation problem.
Casado described the transition from hosting software to hosting agents as “the largest and most important” shift in computing yet, drawing a line from the move off physical servers to virtual machines, then to containers, and now to agent-native compute.
The timing is driven by an infrastructure bottleneck that has received less attention than the GPU crunch. The surge in agent deployments has created what Casado called “a massive CPU shortage alongside the model-induced GPU shortage.” Agents consume significant ordinary compute for tasks like browsing, file operations, and code execution, and the existing abstraction layers add overhead that compounds at scale.
Parenting as a Design Pattern
Dai, a father of two, frames agent management with a parenting analogy. Just as parents childproof the house and keep the credit card out of reach, organizations deploying agents must limit which files an agent can access and how much it can spend, according to TNW. Runta wraps agents in isolated environments with policy-enforced guardrails so a rogue agent cannot run up a bill or compromise a system.
The approach sits at the intersection of two problems enterprise buyers face simultaneously: agents need more compute freedom to be useful, but that freedom creates risk without proper containment. Runta’s bet is that solving both at the infrastructure layer, rather than bolting governance onto existing cloud primitives, produces better outcomes for both performance and security.
Where Runta Fits in the Agent Infrastructure Stack
The $20 million round lands during a period of heavy capital deployment into agent control infrastructure. In the past week alone, Oak raised $60 million for agent identity governance, Valarian closed $50 million for sovereign AI control planes, and Anthropic launched a $1.5 billion implementation joint venture with Blackstone. Runta’s execution layer focus is distinct from these governance and identity plays, targeting the compute substrate rather than the policy or permissions layer above it.
Casado, Yoko Li, and Guido Appenzeller co-authored the a16z investment announcement. The firm has known Dai for nearly a decade, according to the post, dating back to his Cloudflare and Kong tenure.