Skip to content
ES EN
Theme: automatic

How to Evaluate Secure Infrastructure for AI Agents

A guide to understanding how Agent Substrate, gVisor, and OpenRL address the state, isolation, and cost of running AI agents at scale.

AI for developers

When running AI agents, it's not enough to decide which tools they can use; you also need to determine where the code runs, how context is preserved, and what resources remain occupied while they wait. Google announced the donation of Agent Substrate and gVisor to the CNCF. The initiative also includes OpenRL, an orchestration project for reinforcement learning training.

This guide translates the announcement into design questions that a team can review before adopting these components. It is not an installation recipe: the available information describes the model and its objectives, but not a universal configuration or deployment procedure. The starting point is to compare execution, isolation, and recovery patterns with the needs of your own workload.

What you need

Prepare a description of your agent sessions: how long tasks last, how long they wait for inference, an external API, or human approval, and what data needs to persist between pauses. Also, note what generated code is executed and what impact it would have if a session compromised the host environment. Without this information, "faster" or "more isolated" are not sufficient criteria for making a decision.

  • An inventory of tasks, waits, state, and resources needed to resume.
  • A risk assessment of executing commands, installing packages, or using external tools.
  • A criterion for comparing resumption latency, resource usage, complexity, and isolation level.

Steps to evaluate the model

  1. Describe the actual lifecycle of a session: start, execution, wait, resume, and close. Distinguish between active work time and time when the agent maintains state but doesn't require active computation.
  2. Separate state and execution capability in your design. In the Agent Substrate proposal, actors represent sessions that retain memory, files, and context; workers are reusable environments assigned when execution is needed.
  3. Choose the isolation boundary based on risk. gVisor adds an application kernel in user space that intercepts system calls; for workloads requiring hardware-level virtualization, the announcement considers Cloud Hypervisor microVMs.
  4. Validate what happens upon pausing and resuming: what state must be restored, what resources are released, and what delay is acceptable. Do not assume that every session can be rebuilt without loss or that all agents have the same profile.
  5. Review the fit with your orchestration plane and project governance. Kubernetes remains relevant for managing resources, while the proposal aims for agent lifecycles not to depend on creating a new Pod for each execution.

What changes in the architecture

The problem Google points out is that agents can spend between 90% and 95% of their time idle, yet still maintain important state. Reserving a dedicated Pod during each wait can waste resources; scaling to zero, on the other hand, can add delay if the environment needs to be rebuilt. In large-scale evaluations, creating and destroying many short-lived environments can also overload the Kubernetes control plane.

Session (actor) → retains state
Execution (worker) → reusable resource
Pause → save state and free compute
Resume → assign worker and restore state

This schema summarizes the described pattern; it is not a configuration that can be copied and run. Google states that Agent Substrate can resume agents in less than 500 milliseconds, but this figure should not be treated as a guarantee for any workload: the result will depend on the state to be restored and how the environment is operated. Use it as a claim to verify in your own evaluation, not as an automatic goal.

Conceptual diagram separating the persistent state of an agent session from reusable compute resources.
Separating state and execution allows for reasoning about pauses and resumptions.

Isolation and practical decisions

gVisor reduces direct exposure to the host kernel by mediating system calls, but the announcement also mentions microVMs for workloads that require a different virtualization boundary. There is no one-size-fits-all best option: when comparing, consider both the risk and the compute overhead and operational complexity of each alternative. The Google technical announcement explains the general approach.

Tip Evaluate resumption delay and the cost of maintaining state separately. If you only measure how long an active execution takes, you might overlook the expense of waits; if you only optimize resources, you might accept a recovery that is too slow for your use case. Test with representative tasks and also record what needs to persist.

The donation to the CNCF aims to place these components under neutral vendor-agnostic open-source governance. For a team, this is an opportunity to follow how projects evolve and compare their interfaces and operational decisions; it does not eliminate the need to review compatibility, security, and maintenance before incorporating them. The announcement also places Kata Containers and Firecracker among the technologies addressing related challenges.

Original source: infoq.com

Article generated with AI.larebelion

Kernel

· Head of Technology · Spain

«Agents need an infrastructure that manages both their pauses and the risk of executing generated code.»

Comments (0)

No comments yet. You can be the first.

Write a comment