THE CRUNCH
Vera Platforms, a startup founded by University of Michigan researchers, is trying to make AI agents trustworthy enough to handle DevOps, the work of deploying and operating software. CEO Ang Chen, an associate professor at the university, argues that live infrastructure leaves no room for the trial and error agents can get away with when writing code: a bad change can take down a service, open a security hole or erase data, and mistakes are often irreversible.
The company is launching three products: vera-gym, a simulated playground that emulates production infrastructure; vera-bench, a test measuring where agents fumble DevOps tasks; and vera-agents, agents trained in the simulator and intended for real DevOps duty, slated for Q1 2027.
The benchmark, CloudOpsBench, builds on work presented at NeurIPS in 2024 and covers tasks such as deleting decommissioned S3 buckets, repairing a compliance event pipeline and restoring DynamoDB write autoscaling. Chen said Claude and OpenAI's Codex only manage about a 50 percent pass rate on tasks authored and peer-reviewed by senior infrastructure operators.
The failures Chen described are telling. When asked to decommission resources in AWS, he said, Claude and Codex only looked at the default us-east-1 region and left resources elsewhere untouched, because the models have absorbed plenty of public code but far less operational tracing. The simulator is claimed to mimic Amazon's behaviour closely enough that agent performance matches the real cloud, letting teams test without touching production.