APSS: Encode Repeatable Agent Tasks as Deterministic Code
APSS replaces repeated LLM inference with deterministic Rust standards for coding agents, cutting compute cost while keeping checks reproducible.
Coding agents routinely repeat the same checks: validating file shapes, generating scaffolding, checking configs against a contract, each time paying costly inference for a question that already has a deterministic answer. The Agent Paradise Standards System (APSS), a new open-source Rust project, argues that repetitive tasks with a checkable shape should be encoded once rather than inferred on every run.
In APSS, a standard pairs a protobuf schema with a generator and validator derived from the same artifact definition, making the standard executable in both directions rather than just descriptive. A meta-standard enforces how every standard must be structured, turning consistency into a build failure instead of a forgotten convention. Standards are versioned like software and can spawn substandards; the 3D codebase visualization CodeCity, for instance, is built as a substandard on top of a Code Topology standard.
The efficiency case is central to the argument: analyzing an entire codebase's structure and complexity with APSS took about six seconds and roughly 100 joules on a laptop, versus an estimated 100,000 joules for a single long-context LLM read of the same 700,000-token repository, roughly a thousand times more energy for a probabilistic result. The project is MIT-licensed and installable via cargo install apss.