Architecture
System overview, component breakdown, and data flow diagrams.
Risk Scoring
How proposals are scored 0–100 against behavioral patterns and policy rules.
Agent
The Nanobot skill pack — commands, lifecycle, MCP security tools, and Telegram glue.
On-Chain
Squads V4 program, multisig layout, PDAs, and the sponsored transaction model.
Design Principles
- No LLM on the hot path — risk scoring is deterministic and instant. The agent only steps in after a verdict is reached, for communication and ad-hoc analysis.
- Squads V4 is the foundation — battle-tested multisig infrastructure. Sage adds intelligence on top without replacing key management.
- The server owns the second key — Sage holds a member key with
Executepermission. No approved transfer executes without it, making it genuinely agentic. - Sponsored end to end — the server is the fee payer and rent collector for every transaction, so users never need SOL.
- Continuous learning — behavioral patterns are recorded after every executed transfer; the profile improves automatically.
- Conversational, not authoritative — the agent calls endpoints and reports results; it never makes the deterministic decision itself.