Vera develops training tools to improve AI agents’ reliability in DevOps
Vera Platforms, a startup founded by University of Michigan researchers, is developing tools to train and assess AI agents for DevOps work. The aim is to make agents more dependable when changing live cloud infrastructure, where a mistake can disrupt services, expose security weaknesses or erase data.
The company’s products are vera-gym, a simulated infrastructure environment; vera-bench, a test suite; and vera-agents, trained systems intended for deployment. Its CloudOpsBench tasks include deleting retired S3 buckets and restoring DynamoDB autoscaling, and the company says leading models Claude and Codex pass only about half. Vera says its simulator closely matches Amazon’s behaviour, while vera-agents are expected in the first quarter of 2027.
In short
- Vera is building tools to train and test AI agents on DevOps tasks.
- Claude and Codex score about 50 per cent on its benchmark.
- The company expects its trained agents in Q1 2027.
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Tools that train and rigorously test AI agents could make DevOps work safer by exposing weaknesses before agents touch live infrastructure. If simulations reflect real cloud behaviour and benchmarks cover consequential tasks, organisations may be able to automate routine work more reliably, reduce human error and direct expert attention to harder problems.
The case against
Passing simulated tests may not show that an agent can handle the varied, changing conditions of live cloud systems, where a mistaken action can disrupt services, weaken security or destroy data. Benchmarks and claims about simulator accuracy also come from the company developing the products, so organisations may reasonably want independent evidence and strong safeguards before relying on these agents in production.
Explained simply
DevOps teams build and maintain the software and cloud systems that organisations rely on. AI agents are software systems that can carry out tasks, such as changing cloud settings, with less step-by-step direction from a person.
Cloud infrastructure can store data, run services and control who can access them. Changes made in the wrong way may interrupt services, weaken security or delete data, so organisations need ways to check that an agent can handle these tasks safely and reliably.
Vera Platforms is a startup founded by University of Michigan researchers. Its tools are designed to train and assess agents in simulated infrastructure, and to test their performance on practical tasks before they are used in real systems.
AI Art Business Celebrity Companies Culture Entertainment Technology
Read the full article at the source →
Originally published by The Register as “Startup wants to make AI agents safe for DevOps duty”.