Kastor emerges as a declarative toolchain for defining and managing AI agents, aiming to provide a vendor-neutral, versionable source of truth. Unlike imperative approaches common in frameworks like Lang Graph or platform UIs, Kastor uses HCL (HashiCorp Configuration Language) to specify agents, tools, and prompts.

The Go toolchain offers two primary functions: `kastor build` generates runnable projects for target frameworks, while `kastor plan` and `kastor apply` manage agents as long-lived resources on hosted platforms, complete with state, diffs, and drift detection. This approach mirrors Terraform's infrastructure management model for AI agent development.

Currently a proof of concept, Kastor can parse HCL specs, validate references, and build Lang Graph projects. Future plans for version 0 include `plan/apply` capabilities, local state management, drift detection, and deployment to cloud platforms like AWS and Azure. The project offers installation via Homebrew, a script, or direct Go installation.