Vera: Building AI Agents to Run Infrastructure

Today, we are excited to announce Vera’s launch from stealth! We build evals, gyms, and agents for infrastructure operations, pushing the frontier of AI from writing code to running high-stakes infrastructure. Coding agents now excel at software engineering, but provisioning, configuring, and repairing production infrastructure is a different job, and even frontier agents aren’t ready for it. We offer three products: vera-gym is a safe playground that emulates production infrastructure in high fidelity, vera-bench measures where agents fall short, and vera-agents, trained in the gyms, graduate as seasoned DevOps engineers.

Why coding agents can’t run infra (yet)

Coding agents are powerful because they train on enormous amounts of code, and code semantics generalize well: a Python quicksort behaves the same everywhere. Although no agent is 100% reliable, mistakes in everyday coding are cheap because agents can safely run tests against their work. DevOps breaks each of these assumptions.

Opaque platforms. DevOps works on infrastructure platforms whose internal behaviors are entirely opaque; agents only have access to their APIs and documentation, not their internal implementations. Even a single cloud exposes thousands of APIs, with dozens of options just to create a virtual machine. Every infrastructure is also different (AWS vs. Azure vs. GCP vs. OCI), and the API surface keeps shifting over time.

Data scarcity. DevOps programs and traces live inside enterprises, locked away from even the most powerful coding models. They can’t be scraped at scale for training. Compounding this scarcity, every enterprise’s deployment is bespoke, and much of the knowledge about a deployment is tribal. What an agent learns about one platform doesn’t simply carry over to the next.

Safety-critical. Live infrastructure is unforgiving. An incorrect change can take down a service, erase data, or introduce a security vulnerability. These mistakes are costly and often irreversible, and there’s no safe way to experiment. Running code against production just to see what happens isn’t an option.

It’s no accident that software engineer and DevOps engineer have always been different jobs!

The path toward infra agents

For production work, intelligence alone isn’t enough: infrastructure demands safety and reliability. Our vision is that agents need two missing building blocks to get there: a gym to practice in and a curriculum to learn from. vera-gyms are high-fidelity infrastructure emulators where agents can provision, break, and repair infrastructure thousands of times, learning from their mistakes without touching anyone’s production systems[1]. Because each gym encodes the platform’s semantics, it can also check an agent’s plan before it acts[2]. vera-bench finds where agents fall short and turns those blind spots into a curriculum of steadily harder tasks[3]. With a high-fidelity playground and a rigorous curriculum, vera-agents become seasoned infrastructure engineers that have practiced the job, been tested on it, and are guarded by formal checks. We believe this roadmap is the only route to agents we can trust with production. This is what we’re building.

vera-gym: a playground for infrastructure agents

The gym is where agents learn the job. Each vera-gym is a high-fidelity emulator of a real infrastructure platform, faithful enough that an agent’s actions behave the way they would in a real deployment[1]. To construct the gyms, our engine reads platform documentation and real operational traces and synthesizes the emulator automatically, using neurosymbolic program synthesis built on years of our team’s research. The work is led by Raymond Sun, Shaurya Narwal, and Caedmon Swanson, together with research scientist Dr. Xinyu Wang, a leading expert in program synthesis and formal methods.

vera-gym is more than a training ground: it is an executable specification of the platform, with formal guarantees built in. Agents can try a change in the gym before acting and adjust their plans, and the gym can be machine-checked and mined for the rules every safe operation must follow[2]. In effect, we uncover the laws of physics for infrastructure and turn them into runtime guardrails for agents.

vera-bench: benchmarking infrastructure agents

Even the best agents today struggle to be good DevOps engineers, as our benchmark shows (cloudopsbench.ai). vera-bench grew out of IaC-Eval, one of the first benchmarks in agentic DevOps, which we published in the NeurIPS Datasets and Benchmarks track[3]. We have since rebuilt it as a frontier benchmark for agents, shaped by input from many senior DevOps engineers.

The benchmark shows us why agents fail. For instance:

• Knowledge gap: the agent never learned some aspect of the infra.
• Recall gap: it knows the answer but can’t recall it when it matters.
• Tool-use gap: it doesn’t know how to use the right tool.
• Reasoning failure: it can’t work through a complex setup.

…and many more failure modes specific to DevOps work.

Agents fail in prod in their own distinctive ways. Some lack knowledge of how a specific cloud resource behaves; others skip runtime checks[2] or ignore security best practices. vera-bench tells these failure modes apart and measures each one precisely, so every improvement targets a specific weakness. The effort is led by benchmark lead Antonio Mejia, in close collaboration with our research scientist Dr. Quinn Burke, who brings deep expertise in cloud security and reliability.

vera-agents: safe and reliable infra agents

vera-agents learn through a rigorous curriculum and derive their guarantees from the gym. That sets them apart from today’s makeshift approaches, such as repurposing a coding agent for DevOps by handing it documentation – which is rich in information but gives the agent no precise, checkable space to learn from. Our gym grounds every action, lets agents learn infrastructure operations safely, and keeps their actions within the “laws” each platform must obey. Our earlier work built agents for specific DevOps tasks. NSync, developed with AWS researchers, reconciles infrastructure-as-code with configuration drift in the cloud[4]. Lilac translates an existing cloud deployment into infrastructure-as-code configurations[5]. We’re now drawing on that experience to build general-purpose DevOps agents through post-training. Our agent development team works closely with our scientific advisor, Dr. Andrew Barto, an AI pioneer who shared the ACM Turing Award for laying the foundations of reinforcement learning (RL).

Who we are

Our team brings experience from NVIDIA, Microsoft, Meta, Datadog, Airbnb, and Cruise, and includes professors from University of Michigan, University of Arizona, and UMass Amherst. Together we’ve published more than 100 papers on neurosymbolic AI, cloud computing, and AIOps at leading venues, including ICML, NeurIPS, ICLR, SOSP, SIGCOMM, NSDI, ASPLOS, POPL, PLDI, and USENIX Security. We couldn’t be more excited to turn what we’ve learned at the research frontier into tools the world will use.

References

1. A Case for Learned Cloud Emulators. HotNets 2025.
2. Unearthing Semantic Checks for Cloud Infrastructure-as-Code Programs. SOSP 2024.
3. IaC-Eval: A Code Generation Benchmark for Cloud Infrastructure-as-Code Programs. NeurIPS 2024 (Datasets and Benchmarks Track).
4. Automated Cloud Infrastructure-as-Code Reconciliation with AI Agents. ISSTA 2026.
5. Lilac. IEEE AIOps 2025.