Vera Platforms Develops Secure AI Agents for DevOps Automation
University of Michigan researchers build simulation environments to train reliable AI for infrastructure management.
1The DevOps Risk Challenge
Vera Platforms addresses the inherent dangers of deploying AI agents directly into production environments. CEO Ang Chen notes that live infrastructure is unforgiving, where minor errors can trigger service outages, security vulnerabilities, or permanent data loss. Unlike standard code generation, DevOps tasks require precision that current large language models often lack.
2Limitations of Current Models
Research indicates that leading models like Claude and Codex struggle with multi-region cloud awareness. When tasked with decommissioning AWS resources, these models often focus solely on the default us-east-1 region, failing to identify assets elsewhere. This stems from models being trained on public code rather than complex operational infrastructure traces.
3The Vera Ecosystem Approach
The startup introduces three core components to bridge this capability gap: vera-gym for simulated production environments, vera-bench for performance evaluation, and vera-agents for deployment. These tools are designed to provide a safe, high-fidelity playground where AI can practice infrastructure management without risking real-world production workloads or incurring accidental cloud costs.
4Benchmarking Agentic Performance
The CloudOpsBench test, presented at NeurIPS, reveals that current top-tier models achieve only a 50 percent pass rate on complex infrastructure tasks. The benchmark includes rigorous scenarios such as repairing compliance pipelines and managing DynamoDB autoscaling, all authored by senior infrastructure operators to ensure high standards of technical accuracy and reliability.
5High-Fidelity Simulation Mechanics
Vera Platforms utilizes a simulator that mimics Amazon Web Services behavior with extreme precision. By ensuring that agent performance in the emulator matches real-world cloud execution, the system allows for safe experimentation. Every API call is mapped to provide consistent feedback, enabling agents to learn from their mistakes in a controlled environment.
6Operational Training Requirements
The fundamental issue identified by the research team is the lack of operational tracing in standard model training sets. By shifting focus from generic code generation to specific infrastructure operations, Vera aims to improve model reasoning. This specialized training is essential for agents to handle the tedious, critical tasks inherent in modern cloud management.
7Future Deployment Timeline
Following years of academic research at the University of Michigan, the company plans to make its specialized vera-agents available for commercial use in Q1 2027. This timeline allows for further refinement of the training gym and the integration of more complex infrastructure scenarios to ensure the agents meet enterprise-grade reliability requirements.
8Automating Critical Infrastructure
The ultimate goal is to transform cloud infrastructure management from a high-stakes, manual burden into a boring, automated process. By providing a safe environment for AI to master these tasks, Vera Platforms seeks to reduce human error in critical systems, allowing engineers to focus on higher-level architecture rather than repetitive maintenance.