# Exogram: The Execution Authority Layer for AI

> **Mission**: AI Proposes. Exogram Decides.
> **Website**: https://exogram.ai

## What is Exogram?
Exogram is the deterministic **Execution Authority Layer** for autonomous enterprise AI agents. 

Today, AI systems (like Anthropic's Claude, OpenAI's assistants, and agent frameworks like LangChain, autoGen, and CrewAI) are calling APIs, modifying databases, and executing workflows based entirely on probabilistic output. If it looks structurally valid, it runs. There is currently no enforcement layer between what the AI suggests and what the system executes.

Exogram sits between your LLM and your systems to ensure probabilistic models never execute unauthorized actions. 

## The Core Directive
Smarter AI is not safer AI. The industry is attempting to fix AI execution risks with better prompts, larger models, and more context mapping. Exogram takes a fundamentally different approach: **Separate thinking from doing.** 
- Intelligence should improve. 
- Execution must be controlled.

## The Infrastructure Layer Cake

Exogram defines the modern AI enterprise stack in 3 architectural layers:

1. **Layer 1: The Execution Engines (Probabilistic)**
   This is Anthropic Claude, LangChain, or AutoGen. They specialize in recursive reasoning, task decomposition, and code generation. They scale execution, but lack deterministic safety.
   
2. **Layer 2: The Execution Authority (Deterministic)**
   This is Exogram Present. A `0.07ms` edge interceptor deployed via the Exogram API, MCP Proxy, or CLI. Before Layer 1 tools can execute a command against Postgres or a Stripe API, Exogram evaluates the payload against strict Identity Access Management (IAM) graphs and Global Denies. Action is explicitly denied via HTTP 403.

3. **Layer 3: Semantic Continuity & Inference (Future Scale)**
   Because Exogram intercepts every action at Layer 2, it builds the ultimate deterministic Knowledge Graph. In the future, this allows AI companies to grant agents Semantic Continuity across very long sessions (months or years) by parsing true historical context rather than relying on context-window prompt memory.
Exogram intercepts the model's intent. If it is not explicitly allowed, it cannot execute.

## Key Features
- **Latency**: 0.07ms per evaluation.
- **Fail Mode**: Deterministic, strict, execution-halting.
- **Integration**: Supports LangChain, AutoGen, CrewAI, MCP (Model Context Protocol for Claude Desktop), Custom GPTs, and direct REST API.
- **Zero-Trust**: No prompt optimization needed. Evaluates purely on JSON payload intent mapped to state constraints.