Soar, ACT-R & blackboards
Shared state, specialized processes, procedural learning, and Soar’s chunking already establish key precedents.
Concept Paper v0.1 — September 2026
We Gave AI Language.
Maybe It Needs
Machine Code.
A systems hypothesis: agents may preserve performance while reducing autoregressive reasoning by using machine-facing semantic decisions, typed state, and governed reusable fast paths.
This work presents an architectural hypothesis and proposed experimental framework. It does not claim experimental validation.
The central problem
Natural language as a universal bus
Contemporary LLM agents often serialize internal coordination and reasoning through natural language. A model plans, narrates a tool call, interprets textual results, restates state, and repeats.
Language remains indispensable at the human interface. But using it as the default internal computational protocol is an engineering choice—not an axiom.
state → LLM → reasoning text → tool → textual observation → LLM → …
A limited analogy
From logic gates to semantic gates
Logic gates map signals to deterministic outputs. A semantic gate would map state to a constrained, learned probability distribution—context-dependent, fallible, but directly usable by software.
FIG. 01 — The computing stack
STATE→SEMANTIC DECISION→RELEVANT .94 / RISKY .08 / NOVEL .76
A speculative instruction set
Cognitive opcodes
These names are illustrative, not a proposed standard. The research question is which operations are general, separable, calibratable, and composable enough to justify becoming primitives.
OBSERVEATTENDRECOGNIZEABSTRACTRECALLPREDICTHYPOTHESIZEVERIFYASSESSSELECTACTESCALATESTOP
FIG. 02 — A cognitive opcode
Machine-facing coordination
Cognitive Intermediate Representation
Verbose natural-language reasoning
“The test failed because the expected key was 9000 but the loop only produced values through 8999. This looks like an off-by-one error.”
Compact machine-oriented representation
OBS[test=T17,status=FAIL] EXPECT[key=9000] OBSERVED[max_key=8999] HYP[off_by_one=.97] NEXT[inspect:loop_boundary]
A production implementation could use typed objects, latent state, or a hybrid. The point is not the syntax: human-readable prose is not required at every internal boundary.
An event-driven graph around shared state
Cognitive Runtime
Intent, relevance, novelty, risk, memory matches, progress, tool requirements, uncertainty, and language can be evaluated in parallel. Components may be deterministic code, learned classifiers, vision encoders, retrieval systems, specialist models, or a general reasoner.
FIG. 03 — The graph, not any single model, becomes the intelligent system.
Shared state
Goals, entities, evidence, memory references, progress, and uncertainty.
Probabilistic control
Confidence, entropy, abstention, and calibration drive escalation and policy.
Conditional compute
Invoke expensive language reasoning only when the state actually demands it.
System Two can teach System One
Governed cognitive JIT compilation
UNKNOWN TASK→EXPENSIVE REASONING→RECURRING PATTERN→CANDIDATE FAST PATH→VALIDATION→PROMOTION→INEXPENSIVE EXECUTION
This is not merely caching. A promoted fast path must have defined validity conditions, calibration behavior, versioning, test coverage, monitoring, and a reversible rollback path.
FIG. 04 — Compilation is a governed lifecycle, not a shortcut around evaluation.
What is actually being proposed
A narrow, defensible synthesis
Individual ingredients have substantial prior art. This proposal does not claim to invent modular cognition, latent reasoning, cognitive architectures, or compilation of deliberation. Its proposed research direction is their combination:
SEMANTIC GATES→COGNITIVE OPCODES→COGNITIVE IR→HETEROGENEOUS RUNTIME→PROBABILISTIC CONTROL→GOVERNED COMPILATION
Prior art
The idea has ancestry
Neural program systems
Neural Programmer-Interpreters and Neural Module Networks show reusable, compositional computation.
Hierarchical RL / options
Temporally extended actions with initiation conditions and termination are direct precedents for abstraction over action.
MRKL, ReAct & Toolformer
LLMs as components within larger systems, able to reason, act, and use specialized external capabilities.
DSPy & LMQL
Programming models and compilation for LM pipelines—not merely ad-hoc prompt strings.
Voyager
Experience can accumulate as executable, retrievable, and refinable skills.
Coconut & latent reasoning
Reasoning need not be decoded into ordinary language at every intermediate step.
Jev / typed decisions
A timely System-One-style primitive: machine-facing distributions rather than generated strings. Its architectural claims remain vendor claims.
See the supported bibliography and the full paper for precise distinctions and links.
Falsifiable experiment
At matched performance, does it cost less?
Compare a conventional reasoner/tool loop against a cognitive runtime that uses a machine-facing state layer and cheap semantic decisions before and after expensive reasoning. Escalate only for uncertainty, novelty, or policy.
FIG. 05 — A minimal experiment does not require a new foundation model.
TASK SUCCESSREASONING CALLSREASONING TOKENSINPUT TOKENSWALL-CLOCK TIMETOOL CALLSREPEATED / META REASONINGCALIBRATION & ABSTENTIONCOMPUTE / COST WHERE MEASURABLE
Primary source
Read the full concept paper.
The paper contains the complete argument, prior-art discussion, scope note, and references.