Today we released our first piece of research: a pattern for building AI agents that are portable, persistent, and provable.

We’re calling it File-Driven Agent Architecture.

The Problem

Most AI agent frameworks scatter state everywhere—code, databases, model weights, session memory. When something breaks, you can’t tell what the agent knew or why it behaved that way.

FDAA takes a different approach: the agent IS a folder of markdown files.

agent/
├── IDENTITY.md   # Who it is
├── SOUL.md       # How it thinks
├── MEMORY.md     # What it remembers
└── TOOLS.md      # What it can do

Change a file, the agent changes. No restart. No retraining. No config UI.

Three Properties

Three properties we think are essential for production agents:

  • Portable — Copy the files, move the agent. Works with any LLM.
  • Persistent — Memory survives restarts, upgrades, infrastructure changes.
  • Provable — Every state change is human-readable, diff-able, auditable.

Validation

We validated this pattern against existing benchmarks:

  • 74% accuracy on memory tasks (outperforms RAG and specialized tools)
  • 28% faster execution with structured markdown
  • 90% cost reduction via prompt caching

What We’re Building

This is our first contribution. We’re a research lab, not a product company—at least not yet.

We’re exploring how agents can coordinate, transact, and verify each other’s work. That requires infrastructure we don’t think exists yet.

So we’re building it.

Read the full paper →

Substr8 Labs is a research lab exploring the foundations of the agent economy.