How Runta is Building the Execution Layer for Agents

Runta, building an agent execution layer for token efficiency, secret protection, and governed access, led by Guanlan Dai, has raised $20M in seed funding led by Martin Casado at Andreessen Horowitz, with participation from Jeff Dean, Fei-Fei Li, Ali Ghodsi, Ram Shriram, and Thomas Wolf.

The foundation of modern software infrastructure has long relied on systems that enforce security, reliability, and governance before applications execute. Technologies such as edge proxies, caching, web application firewalls, authentication, authorization, rate limiting, and circuit breaking all operate on a core assumption: software follows predefined logic and executes only the paths developers have explicitly written. This predictable behavior has enabled organizations to secure and manage applications by controlling what enters and exits their systems.

AI agents operate differently from traditional software by making decisions dynamically during execution rather than following predefined paths. Instructions embedded in incoming data can influence model behavior, tool usage, or credential access in ways that appear legitimate. As a result, conventional security approaches that focus on monitoring inputs and outputs are no longer sufficient. Effective governance must occur at the execution layer, where an agent’s decisions translate into real-world actions.

What Is Runta?

Runta is an execution layer designed to govern AI agents as they operate. Rather than requiring changes to existing agents or frameworks, it provides a controlled runtime environment that manages operating system, network, file system, and credential access while agents execute.

The platform enforces predefined policies at the moment actions occur, ensuring that every file access, tool invocation, network request, and credential usage is governed in real time rather than relying on an agent’s stated intent.

Runta’s capabilities focus on three key areas:

• Resource management: Monitors process activity, suspends idle agents, and provides visibility into compute and token usage to help control operational costs.

• Access control: Applies task-specific permissions, issuing credentials just in time and restricting network and resource access based on predefined policies.

• Observability and audit: Records system calls, network activity, file operations, credential usage, and policy enforcement events, creating a comprehensive execution record for auditing, debugging, and recovery.

The execution environment launches in milliseconds and dynamically scales based on an agent’s workload, providing governance with minimal performance overhead. Rather than restricting capabilities, this lightweight execution layer enables AI agents to handle real-world tasks safely and reliably by enforcing control where it matters most.

What's Next?

Runta is building an execution layer designed for production AI agents, supported by a $20 million seed funding round. The investment will accelerate development of infrastructure that enables organizations to deploy AI agents safely in real-world environments.

As businesses seek to automate increasingly complex workflows, widespread adoption depends on providing agents with controlled access to systems, data, and tools. By enforcing governance, visibility, and policy at runtime, Runta aims to give organizations the confidence to expand the scope of work AI agents can perform securely and reliably.