Vera

Agents got good at writing codeRunning infrastructure is next

Vera builds the gyms, the benchmarks and the agents for infrastructure operations

Illustrative terminal session. A project is initialized and the emulator starts a gym across Amazon Web Services, Google Cloud and Microsoft Azure. The benchmark scores the agent at 41 of 60 tasks, the agent trains in the gym on its 19 failures, and the benchmark then scores it at 58 of 60: a trained agent, ready to use.

Three problems prevent reliable infra ops agentsVera solves that

  1. Problem: Real clouds are risky places to learn

    Vera’s answer: Gyms emulate real clouds for safe practice

  2. Problem: One wrong change can be irreversible

    Vera’s answer: The benchmarks verify the solution an agent creates

  3. Problem: DevOps training data is scarce

    Vera’s answer: We supply real tasks and speed up agent training

From practice to production

  1. Gyms

    vera emulator

    • Amazon Web Services (AWS)
    • Google Cloud
    • Microsoft Azure
    • Oracle Cloud Infrastructure, Coming soon
    • More coming
    1. Run
    2. GradeFAIL 3/4
    3. Reset

    Agents practice on emulated clouds running locally, with no real cloud resources provisioned.

  2. Benchmarks

    vera bench

    CloudOpsBench, top four agents by resolution rate
    RankModel and harness
    01claude-fable-5-1claude-code
    02gpt-6-astracodex
    03gemini-4-argongemini-cli
    04grok-4.7cursor-cli

    Graded on the end state, so any valid solution passes. Evaluates agents (model + harness), not just models.

  3. Agents, In development

    vera agent

    Preview

    vera agent "rotate the expiring TLS

    cert on the api load balancer"

    acmcert expires in 3 days

    acmrequested new certificate

    elbv2attached to listener :443

    ✓ doneapi → 200 OK

    Trained in the gyms, graded by the benchmarks, then put to work.

Research

View all research

Get in touch

Tell us what you’re working on and we’ll get back to you

Enter your name.

Enter an email address we can reply to.

Tell us a little about how we can help.