Most AI browser automation is a liability. You point an agent at a website, hope it does the right thing, and pray it doesn't click "Confirm Purchase" on something you didn't authorize.
We built something different: a governed browser swarm where every agent action passes through a 14-layer security pipeline, and different agents literally see different shortest paths through the web because their mathematical personalities warp the geometry.
The Architecture
User Request > FastAPI Gateway (port 8000) > Governance Membrane (14 layers) > TongueRouter (Dijkstra on Poincare Ball) > Agent Pool > Playwright Wrapper > Training Tap
The Tongue Router
Every URL gets expanded into a 6D vector using Sacred Tongue space - six dimensions weighted by the golden ratio:
class TongueObserver:
TONGUE_WEIGHTS = {
'KO': 1.000, # Intent
'AV': 1.618, # Context
'RU': 2.618, # Policy
'CA': 4.236, # Execution
'UM': 6.854, # Security
'DR': 11.090, # Attestation
}
Agent Personalities
Different agents have different tongue weightings - they see different edge costs in the navigation graph. A scout finds the fastest route. An auditor finds the safest route. Same web, different geometry.
The math: g_ij(x, agent) = (4/(1-|x|^2)^2) * T_ij(agent)
Governance at Every Step
- Semantic antivirus - checks intent against known malicious patterns
- Action validation - confirms the action matches the approved task
- Governance scan - 14-layer pipeline produces ALLOW/DENY/QUARANTINE
- Training tap - every interaction becomes an SFT training pair
14,654 training pairs and counting, all pushed to HuggingFace.
Running It Yourself
git clone https://github.com/issdandavis/SCBE-AETHERMOORE.git
cd SCBE-AETHERMOORE
pip install -r requirements.txt
python -m uvicorn src.api.browser_saas:app --port 8000
Why Open Source?
AI governance is too important to be proprietary. The full codebase: github.com/issdandavis/SCBE-AETHERMOORE
Patent pending (USPTO #63/961,403) - the math is protected, the code is free.
Top comments (0)