Cloud-native infrastructure is defined almost entirely through configuration: Kubernetes manifests, Helm charts, and Infrastructure-as-Code (IaC) such as Terraform. As these ecosystems grow in scale and evolve rapidly, writing correct configuration has become one of the most error-prone and labor-intensive parts of operating modern systems, and misconfiguration remains a leading cause of outages and security incidents. Historically, practitioners have relied on hand-written templates, schema validators, linters, and policy-as-code engines. However, schema validators check only syntax and field types, policy rules are brittle and costly to maintain, and templates cannot adapt to new requirements. None of these methods can translate a natural-language operational intent into a working configuration, and all of them struggle with constant API and provider version drift, where a configuration that was valid yesterday silently breaks against today's schema.
To address these limitations, Large Language Models (LLMs) and coding agents have been applied to generate and repair cloud-native configurations directly from natural-language requirements and incident reports. Agents such as Claude Code and Codex can inspect cluster state, write manifests and Terraform modules, invoke validators, and iterate on their own output. However, applying them to production infrastructure faces hurdles: models hallucinate fields and API versions, produce configurations that pass syntax checks but cannot be deployed or do not match the intended behavior, and are difficult to evaluate fairly because benchmarks can leak reference solutions to the agent. Despite these challenges, LLM-driven configuration generation is a rapidly expanding field, generally categorized into three research directions: prompt-based generation, verifier-in-the-loop repair, and memory- and retrieval-augmented agents.
We are working on making LLM agents reliable, self-improving engineers of cloud-native configuration. Our research focuses on autonomous research agents that generate configurations, verify them against real execution environments, and recursively improve by writing verified experience back into a typed, validity-scoped memory, so that every validated success or failure informs later generations without carrying stale knowledge across platform versions. For example, we have developed a verifier-guided agent framework that couples deterministic checks (schema validation, Terraform planning against a mocked cloud, and dynamic evaluation on live Kubernetes clusters) with an intent review step, and evaluates Claude Code and Codex on Kubernetes incident repair and the IaC-Eval benchmark under leakage-audited conditions. In our experiments, the verification loop alone raised the share of Codex-generated Terraform that plans successfully from 61% to 98%. This methodology lets the system converge toward correct, deployable configurations while giving operators auditable evidence of why each configuration can be trusted.