Skip to content
pnxt

Introduction to pnxt

pnxt is a research project designing a net-new programming paradigm built exclusively for LLMs and AI agents, moving beyond human-readable legacy syntax toward structured, verifiable, graph-based program representations.

The Problem

Traditional programming languages are optimized for human cognition — visual hierarchy, short-term memory, lexical parsing. LLMs, however, excel at structural data manipulation (JSON/graphs) but struggle with implicit control flow and loop-state tracking.

Current approaches try to make LLMs better at writing human-centric code. pnxt asks a different question: what if we designed programming from the ground up for how LLMs actually think?

The Vision

pnxt designs an execution environment where LLMs orchestrate logic graphs rather than generate syntax. In the Agent-Native Programming (ANP) paradigm, AI agents are first-class entities with identity, memory, state, and tools — not just coding assistants bolted onto existing workflows.

Core Principles

Incrementalism

Every component is designed for phased adoption. Start small, grow organically as trust and tooling mature.

Structural Safety

Correct behavior is easy by design. Incorrect behavior is structurally difficult — not relying on discipline.

Explicit Over Implicit

Side effects declared, capabilities negotiated, trust measured. Nothing important is left to convention.

Human Partnership

Agents as colleagues with graduated trust. Humans retain authority over consequential decisions.

What Makes ANP Different?

Unlike conventional agent frameworks that bolt LLM capabilities onto existing paradigms, pnxt proposes genuinely novel contributions:

  • Persistent evolving memory — agents learn and grow across sessions
  • Natural language as interface — not just for prompts, but as architectural lingua franca
  • Negotiated capability contracts — agents discover and negotiate what they can do
  • Graduated trust as architectural concern — trust is measured, not assumed
  • VPIR reasoning chains — verifiable, graph-based program representations with formal execution
  • Z3 formal verification — 21 SMT-verified properties including noninterference, DPN liveness, and graph pre-verification
  • HoTT categorical structure — code as categorical objects with univalence and transport
  • Neurosymbolic bridge — P-ASP confidence scoring and Active Inference graph patching
  • Self-hosting vision — the pipeline describes, validates, and executes itself as VPIR

Getting Started

Quick Start Get up and running in 5 minutes — try VPIR, Bridge Grammar, Z3, and the full pipeline.
Three Pillars of ANP Understand the foundational pillars: ACI, Semantic Memory, and Natural Language.
Research Overview Explore the complete Phase 3 research covering all aspects of ANP.
Project Status See where pnxt is today and what's coming next.
Theoretical Foundations Dive into HoTT, VPIR, DPN, LLMbda Calculus, and SMT solvers.