What if everything about your AI agent — its identity, memory, capabilities, and personality — lived in a single folder of markdown files?
That’s FDAA. File-Driven Agent Architecture.
The Problem With Today’s Agents
Most AI agents today are ghosts. They exist only in RAM. When the session ends, they vanish. Want to know what happened yesterday? Good luck parsing logs.
Worse: you can’t prove what an agent did. You can’t audit its decisions. You can’t verify its identity. You just… trust it.
That’s not good enough for production. That’s not good enough for enterprises. That’s not good enough for regulated industries.
The Insight: Files Are Forever
Here’s the key insight: the intelligence is in the model, but the identity is in the files.
An LLM is a general-purpose reasoning engine. It doesn’t know who it is until you tell it. So tell it — in files:
my-agent/
├── SOUL.md # Who am I? Personality, values, voice
├── IDENTITY.md # Name, role, constraints
├── MEMORY.md # Long-term knowledge
├── memory/
│ └── 2026-02-15.md # Daily logs
└── TOOLS.md # What can I use?
Now your agent has a birth certificate. You can:
- Version it in Git
- Diff changes over time
- Roll back to previous states
- Prove exactly what configuration was running
Why This Matters
Portability: Move your agent between providers. Claude today, GPT tomorrow, local Llama next week. The identity travels with the files.
Persistence: Agents remember across sessions. Not because of some black-box vector database, but because you can read the files yourself.
Provability: Every decision traces back to a file. Auditors can verify. Compliance teams can sign off. Lawyers can sleep at night.
Composability: Agents can spawn sub-agents. Each sub-agent gets its own folder, its own identity, its own audit trail.
The Bigger Picture
FDAA is the foundation layer of what we’re building at Substr8 Labs. On top of it:
- FDAA Verification Pipeline — Prove that skills are safe before execution
- Agent Capability Control — Prove that agents only do what they’re allowed to do
Together, these create something new: provable agents. Not “trust me bro” — cryptographic receipts.
Read the Full Spec
The complete FDAA specification is available on GitHub:
Or dive into the reference implementation:
💻 fdaa-cli — Verification pipeline in Python
This is Part 1 of our research series on provable agent infrastructure.