AI & Agents

Vera Platforms Wants to Make AI Agents Safe Enough for DevOps

Leave a comment

Letting an agent write code is one thing. Letting it touch live infrastructure is another.

The pitch

Vera Platforms, a startup out of the University of Michigan, says “live infrastructure is unforgiving.” CEO Ang Chen told The Register that leading models stumble on ops work. One example: when asked to decommission AWS resources, Claude and Codex only looked in the default us-east-1 region and missed the rest. The company has three products: vera-gym, a simulator that mimics production cloud behavior; vera-bench, including a CloudOpsBench test where Chen says Claude and Codex pass only about 50% of the time; and vera-agents, trained in that gym for DevOps jobs.

My take

The region example is exactly the kind of quiet mistake that bites. Until agents prove themselves, the safe pattern is still a sandbox or simulator first, a plan you review, and a human hitting the button on anything destructive.

God Bless, and Use the Source!

References & Further Reading

Subscribe (RSS) Follow AI & Agents via RSS

Leave a note

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.