Show HN: An LLM agent that emits typed intent (github.com)

A new reference implementation for large language model (LLM) agents has been announced, featuring a pattern of intent-execution separation. This project, dubbed SIF (Structured Intent Format), allows LLMs to emit typed intents using a defined vocabulary without direct interaction with any underlying storage or database systems. Instead, a deterministic layer is responsible for validating, translating, scoping, and executing these commands. This separation aims to enhance security by preventing direct manipulation of the substrate by the LLM and provides backend neutrality, allowing the same ontology to work with various data sources—such as SQL databases or REST services—without altering the LLM's operations. The significance of this project lies in its innovative approach to employing LLMs as intent routers rather than executors, addressing common issues of trust and reliability when LLMs interact with complex business logic. The framework includes a coherent structure for generating business applications using three declarative artifacts: an ontology, a set of business rules, and an agent prompt. By focusing on a smaller vocabulary and predefined operations, the architecture minimizes the risk of hallucinations and misinterpretations, ultimately providing a more controlled and trustworthy environment for automated decision-making in domains such as legal and veterinary applications. The design challenges the norm by proposing that useful applications can emerge from structured ontologies rather than traditional imperative programming.