Grenat is an open-source compiled programming language designed specifically for building AI agent systems. Its Ruby-like syntax and Rust-oriented implementation provide first-class constructs for typed prompts, tools, supervised actor agents, budgets, durable workflows, model evaluation, and integrations such as HTTP, email, databases, MCP, SSH, and embeddings. Prompt return types become JSON schemas, while model outputs are treated as untrusted until they are validated, approved by a human, or explicitly trusted. The compiler checks types, effects, taint, and secret handling; the project says an unvalidated model answer reaching a network, file, command, email, or web page produces a compile-time error, and its effect system is intended to turn prompt injection into a compile-time failure. Workflows journal each step so interrupted runs can resume, including after a delayed human response, without billing the same model call twice. The toolchain includes offline tests with mocks and cassettes, model-backed evaluations with an LLM judge, a language server, formatters, application and agent generators, an operations console, encrypted environment-specific credentials, and native builds through Cranelift or LLVM. It supports configured connectors for multiple model providers and separate embedding and transcription services, but the article does not present Grenat as a model provider. The project reports a comparison in which models given only its roughly 4,000-token `llms.txt` reference completed a support-agent task and a weekly release-digest task with 88,000 tokens across both tasks versus 118,000 in Python, while Python used fewer tokens on the digest alone. The same comparison reported 64 versus 237 lines for the support task and 79 versus 207 for the digest. These results came from two runs per task and are presented by the project. The latest release described is v0.1.2; the article also lists phases 12 to 15 on the main branch, covering libraries, email intake, and other production features. Grenat can be installed on supported glibc-based Linux distributions or macOS, used through Docker, or built from Rust sources; Alpine binaries are unsupported.
