Architecture

Overview

Agent Framework (LangGraph / AutoGen / etc)
        │
        ▼
      MCP Interface
        │
        ▼
┌─────────────────────────┐
│  Substr8 Governance     │
│  ├─ FDAA (Agent Spec)   │
│  ├─ ACC (Policy)        │
│  ├─ RIL (Run Lifecycle) │
│  ├─ DCT (Audit Ledger)  │
│  ├─ CIA (Integrity)     │
│  └─ GAM (Memory)        │
└─────────────────────────┘
        │
        ▼
    RunProof Artifact

Components

FDAA — File-Driven Agent Architecture

Defines agent identity as files:

• agent.md — Agent specification
• tools/*.md — Tool definitions
• Produces deterministic hash for versioning

ACC — Agent Capability Control

Policy enforcement:

• What tools can this agent use?
• What data can it access?
• Runtime capability checking

RIL — Run Integrity Layer

Run lifecycle management:

• Run ID generation
• Start/end tracking
• State machine for run phases

DCT — Deterministic Computation Trail

Tamper-evident audit ledger:

• Hash-chained entries
• Every action recorded
• Verifiable offline

CIA — Conversation Integrity Assurance

Request validation:

• Validates tool_use/tool_result pairing
• Repairs broken message sequences
• Produces integrity receipts

GAM — Git-Native Agent Memory

Memory with provenance:

• Semantic search
• Git-backed versioning
• Typed hints for retrieval

Data Flow

1. Agent starts run
   └─→ RIL generates run ID

2. Agent calls tool
   └─→ ACC checks policy
   └─→ DCT records action
   └─→ CIA validates request

3. Agent writes memory
   └─→ GAM stores with provenance
   └─→ DCT records operation

4. Run completes
   └─→ RIL finalizes run
   └─→ RunProof generated

Skill-Based Architecture

Governance is invoked as tools, not intercepted:

❌ OLD: Client → Proxy → Gateway → LLM
         (proxy in traffic path)

✅ NEW: Client → Gateway → LLM
                   ↓
            Governance Skill
         (invoked when needed)

Benefits:

• No header modification
• Isolated failures
• Lower latency
• Easier debugging