We Gave AI Language. Maybe It Needs Machine Code.
A concept paper exploring machine-facing decisions, shared state and governed fast paths for AI agents.
Does every internal operation need to become language?
An agent writes a plan, describes a tool call, reads a textual result and explains what it will do next. Language serves both as the human interface and as the system’s internal coordination mechanism.
This concept paper asks what happens when we separate those roles. People still communicate with the system through language. Internal components can work with structured state, constrained decisions and explicit uncertainty.
From prose to usable decisions
A component deciding whether a result is relevant does not necessarily need to write a paragraph. It could return a typed decision that software can use directly. The paper calls these components semantic gates.
Combined with a machine-facing intermediate representation — a cognitive IR — this suggests an architecture in which goals, observations, evidence and next steps do not need to be repeatedly serialized into prose.
The system around the model
A cognitive runtime coordinates different components around shared state: deterministic code, retrieval, small specialist models and a general reasoner. Expensive language-based reasoning is invoked when the situation warrants it, including novelty or insufficient confidence.
Recurring solutions can become candidates for faster execution paths. Promotion requires validation, versioning, monitoring and rollback. A previously successful answer alone is insufficient evidence to trust a new route.
A hypothesis to test
This is an architectural hypothesis and proposed experimental framework. It does not claim experimental validation. Its individual ingredients also have substantial prior art: cognitive architectures, modular computation, reusable skills and reasoning outside natural language.
The testable question is whether this combination can preserve task performance while reducing autoregressive computation. Evaluation must compare task success, reasoning calls, tokens, latency, calibration and abstention.
Read the original
This page introduces concept paper v0.1, published in September 2026. The original contains the full argument, scope discussion and bibliography.