AEGIS

AI trading agents run with direct access to exchange keys, so one misread decimal or a blind retry after a timeout can turn into a loss fast. In AEGIS the model never grants permission: every proposal passes a policy layer that sends nothing to the exchange if the data is stale, the same intent repeats, the previous order's outcome is unclear or the amount is over the limit. Every rejection is logged with a reason code.

It comes in two parts: a risk service that judges any agent's proposal, and an autonomous reference agent that applies the same rules to itself. Market data, candles, account and orders are read through Agent Trade Kit, and the digest of a HALT decision was written to the RiskDecisionRegistry contract on Avalanche Fuji.