Research Paper v0.1 Code: Ready on GitHub

An Agent-Level Language for Executable Intent

INTHON is a Python-hosted language layer that allows AI agents to express workflows as compact, deterministic, and auditable programs.

Harsha Vardhan

Chief Researcher

INTHON Research Group • Department of Advanced Computing & Research

HarVa DeepLabs

Try Playground Read Paper (PDF)

Quick Installation

Install the stable package directly from PyPI:

pip install inthon

Abstract

Current autonomous agent architectures communicate actions using natural language, fragile JSON blocks, or boilerplate-heavy Python. While powerful, these approaches suffer from token bloat, execution non-determinism, and lack of rigorous sandboxing.

We introduce INTHON (Intelligent Python), a language layer optimized for the token efficiency of Large Language Models (LLMs). INTHON models workflows as structured agent blocks with strict capability-based policies. By enforcing static validation of tool schemas, sandboxing side effects, and emitting detailed, replayable JSON trace graphs, INTHON guarantees auditability and safety at machine execution speeds.

The Execution Pipeline

INTHON translates source programs to Intermediate Representation (IR), validates capabilities, and evaluates tool transactions inside sandboxed environments.

Lex & Parse

Tokenizes stream & matches syntax using Lark grammar rules.

lexer/, parser/

→

AST Generation

Creates immutable syntax nodes representing instructions.

ast/nodes.py

→

Semantic Analyzer

Resolves variables, validates imports, checks static types.

semantic/analyzer.py

→

Policy & Guard

Checks capability boundaries & registers approval gateways.

policy/engine.py

→

Sandbox Execution

Evaluates IR and handles sandboxed tool interactions.

runtime/sandbox.py

Select an execution stage above to inspect details

Hover over or tap any pipeline stage to view its architectural mechanics, component boundaries, and matching Python module source references.

Interactive Code Playground

Experience INTHON compilation and execution behavior. Select a preset program, run the virtual environment, and audit the live execution logs.

.inth

// Select a program from the tabs above

Audit Trace Terminal

Idle

> System initialized. Ready to execute.

Core Innovations

Token-Optimized Syntax

Drastically reduces the token footprint for agent planning, system instructions, and tool calling definitions. Replaces verbose JSON schemas with concise language declarations.

Capability-Based Sandbox

Fine-grained execution policies. Control internet traffic, local filesystem writes, memory persistence, and package import restrictions directly within the language construct.

Deterministic Replay Audits

Synchronously records variables, state changes, tool transactions, API costs, and evaluation metrics into a cryptographic JSON trace log for complete reproducibility.

Development Roadmap

v0.1

Language Core MVP (Current)

Python-hosted runtime with Lark parser grammar, sandboxed AST-walking interpreter, Pydantic tool registry, and initial Python modules bridge.

v0.2

Developer Experience Expansion

Formatting suite (`inthon fmt`), style linter, active interactive REPL console, Tree-sitter configurations, and Language Server Protocol (LSP) for VS Code.

v0.3

Ecosystem Integration

Autodoc tools mapping OpenAPI endpoints to INTHON modules, vector database memory handlers, and Interactive Notebook execution cells.

v1.0

Stable Production Release

Finalized language specification, security audit, sandboxed micro-containers for remote runtime environments, and public tool library registry.

Cite INTHON

If you use INTHON in your research, please cite our concept paper:

@article{vardhan2026inthon,
  title={INTHON: Agent-Level Programming Language Layer for AI-Native Workflows},
  author={Vardhan, Harsha},
  journal={Department of Advanced Computing \& Research, HarVa DeepLabs},
  year={2026},
  url={https://github.com/harvatechs/inthon}
}