# Vera Platforms Develops Safety Framework for AI-Driven DevOps Automation University of Michigan researchers build simulation environments to train reliable AI agents for infrastructure. - **Source:** [www.theregister.com - Articles](https://brifing.si/source/2c34ea50-d341-4a10-925d-12c012a9ad10) - **Published:** 2026-10-29 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/68626ef1-8a21-48b9-a629-d35ecc9c89ef ## The DevOps Reliability Challenge DevOps environments are unforgiving, as AI-generated code mistakes in production can lead to irreversible data loss or security vulnerabilities. Unlike standard software development, live infrastructure does not allow for experimental errors. Vera Platforms, founded by University of Michigan researchers, aims to bridge this critical gap by creating safer, more reliable agentic workflows. ## Limitations of Current Models Leading models like Claude and OpenAI's Codex often struggle with complex operational tasks. Research indicates these models frequently fail to account for multi-region cloud configurations, such as decommissioning resources outside the default us-east-1 region. This stems from a lack of exposure to operational tracing compared to the vast amounts of public code. ## Vera Platforms Product Suite The startup is launching three core products to address these deficiencies: vera-gym, a high-fidelity production infrastructure emulator; vera-bench, a specialized testing framework for agent performance; and vera-agents, which are models trained specifically for DevOps tasks. These tools collectively aim to provide a controlled environment for training and validating autonomous infrastructure management systems. ## Benchmarking Agentic Performance The CloudOpsBench test, building on 2024 NeurIPS research, evaluates agents on complex tasks like repairing compliance pipelines and restoring DynamoDB autoscaling. Currently, top models achieve only a 50 percent pass rate on these rigorous benchmarks. These tasks are authored and peer-reviewed by senior infrastructure operators to ensure they reflect real-world operational complexities. ## High-Fidelity Simulation Mechanics Vera-gym provides a high-fidelity emulator that mimics Amazon cloud behavior, ensuring that agent performance in the simulator matches real-world cloud deployments. By mapping every API call, the system allows agents to experience identical environments during training and execution. This approach eliminates the need for risky, direct experimentation on live production cloud infrastructure. ## Academic Roots and Methodology Founded by Ang Chen and his colleagues, Vera Platforms leverages academic research to solve the problem of tedious but critical infrastructure work. The team focuses on creating agents that understand operational context rather than just generating code. Their methodology emphasizes rigorous testing protocols to ensure that AI agents can handle production workloads without human intervention. ## Future Deployment Timeline Vera Platforms has scheduled the availability of their specialized vera-agents for Q1 2027. This timeline allows for continued refinement of the simulation environment and the benchmarking suite. By focusing on high-stakes infrastructure tasks, the company aims to transform how organizations manage cloud resources, ultimately making the maintenance of complex systems significantly more predictable. ## Strategic Industry Implications The shift toward agentic DevOps represents a move to automate mundane, high-risk infrastructure tasks. By prioritizing safety and simulation-based training, Vera Platforms addresses the primary barrier to AI adoption in production environments. If successful, this approach could standardize how enterprises deploy autonomous agents to manage critical cloud services, reducing human error significantly. ## Sources covering this story - https://www.theregister.com/ai-and-ml/2026/10/29/startup-wants-to-make-ai-agents-safe-for-devops-duty/5306477 — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._