We are building the layer between
"I should have sold" and "I sold in time."
Most onchain losses are not bad picks. They are bad exits.
DeFi runs on agents now. Lending agents borrow. LP agents rebalance. Trading agents execute. Every category of onchain action has been delegated to autonomous code — except the one action that matters most when things go wrong.
Risk monitoring has always lived in dashboards: tools that surface signals and leave the decision to you. That model fails when the decision window is ninety seconds long and you happen to be asleep.
The next layer of DeFAI is not analytics. It is delegation.
Four commitments we won't walk back.
Build in public.
Rulebook drafts, threshold tuning, known false positives — published before they're polished. If you can argue with it, we can fix it faster.
Promise only what we can ship.
No claimed audits before they exist. No autonomy before it's tuned. The version of Cassandra at launch is intentionally bounded.
Noncustodial is non-negotiable.
Your funds never leave your wallet. The agent's authority is bounded by code, not promised by us. Cassandra cannot be the rug because Cassandra was never given the key.
Determinism over vibes.
Every verdict is reproducible. Same inputs, same call. No model that decides differently on Tuesday than Monday for reasons no one can explain.
Three things converged that didn't exist before.
- [ 01 ]
Base became liquid enough.
Cassandra's exit routes only work when there is a route to exit through. Base now hosts deep DEX liquidity across the long tail.
- [ 02 ]
Onchain identity for agents is real.
Virtuals Protocol gave agents a wallet, a token, and a verifiable address. Cassandra can be onchain-native instead of a hosted SaaS.
- [ 03 ]
The rug taxonomy stabilized.
After two cycles of post-mortems, the precursor signals to most exit-scams and slow rugs are now reproducible. The rulebook is buildable.
We respond.
Found a rule we missed, a threshold we got wrong, or a historical rug we should backtest against? Tell us. Contributions that materially change a rule will be credited in the next version of the rulebook.