# Exogram: The Execution Authority Layer for AI (Full Technical Summary)

> **Website**: https://exogram.ai
> **API Base**: https://api.exogram.ai
> **Throughput**: 137 RPS (0.00% False Negatives)

## The Core Defect: Unchaperoned Agent Execution
Autonomous agents rely on LLMs, which operate entirely on probabilistic next-token prediction. When agents attempt to mutate production state (e.g., creating a user, deleting a billing record, sending emails via an internal tool), probability introduces catastrophic execution risks.

Foundational models (like Claude, OpenAI, Gemini) and orchestration frameworks (LangChain, AutoGen, CrewAI) focus on "Capability." They generate plans and execute tool calls. They lack "Authority." A hallucinated intent that looks structurally valid as JSON will be executed by an unguarded system. Smarter AI is not safer AI.

## The Exogram Solution: The Infrastructure Layer Cake
Exogram replaces probabilistic guessing with mathematical certainty. We define the modern AI enterprise stack in 3 architectural layers:

### 1. Layer 1: The Execution Engines (Probabilistic)
Foundational models (Claude, OpenAI) and orchestrators (LangChain, AutoGen) operate here. They specialize in recursive reasoning, task decomposition, and code generation. They scale execution, but lack deterministic safety. If a hallucinated intent looks structurally valid as JSON, they will blindly execute it.

### 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. Just as Auth0 governs human permissions, Exogram governs what autonomous agents are allowed to do. 
- The model proposes the action.
- Exogram evaluates the proposed action against an identity access management (IAM) graph and Global Denies.
- The system executes the action ONLY if Exogram explicitly allows it via HTTP 200. Otherwise, it hard-rejects with HTTP 403 Forbidden.

### 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). Instead of relying purely on an LLM's context window memory, agents will query Exogram's historical ledger to gain true contextual inference and prevent cascading failures based on deep structural meaning.
### 2. Deterministic Semantic Conflict Resolution
If an agent proposes a new fact (e.g., "Delete user record X") that lacks proper preceding context (e.g., "Manager approved deletion of X"), the typical orchestration framework executes it anyway.
**Exogram's Fix**: Exogram evaluates the collision against a Supabase Relational Edge Graph via Python logic gates (0.07ms evaluation). If the required permission nodes are missing, Exogram instantly terminates the execution with an explicit `HTTP 403 Forbidden: Missing execution authority`.

### 3. Execution Idempotency & Circular Graph Prevention
Agents often experience network timeouts and retry identical execution tokens (The Double-Spend problem). Furthermore, multi-agent frameworks risk infinite loops when resolving hierarchical conflicts.
**Exogram's Fix**: Exogram acts as an atomic execution proxy.
- It intercepts every tool-use action, assesses meaning, and issues a cryptographic `evaluation_id` mapped to an `idempotency_key`. 
- Over the next 500ms, if the exact same payload hits the endpoint during a retry, a cryptographic lock triggers an `HTTP 409 Conflict: ALREADY_EXECUTED`, mathematically preventing double-spend anomalies.
- A Recursive CTE Postgres trigger monitors all new graph relations and hard-rejects topological loops.

### 4. Legacy Phantom Edge Mitigation
In older AI architectures, semantic search relies on probabilistic proximity. Outdated facts retrieved by an LLM lead to hallucinated logic paths. By acting as the Knowledge Graph Execution Authority, Exogram instantly tags deterministic Node metadata with `{"archived": true}` the exact millisecond a fact supersedes an old one in the Supabase state ledger. This severs deprecated paths before the LLM can ever traverse them.

## Supported Architectures & Integrations
Exogram is an independent execution governance proxy and sits securely above:
- **Anthropic Claude 3.7 Computer Use APIs (MCP)**
- **Meta Llama Tool Calling APIs**
- **LangChain / LangGraph execution loops**
- **CrewAI and Microsoft AutoGen agent systems**

**If your autonomous agents mutate state, Exogram provides the deterministic, zero-trust authorization pipeline you need.**