Heimdall

a flight recorder for autonomous AI agents, on Aurora PostgreSQL

We don't use the database as storage - we use it as the agent control plane: trace store, semantic judge, policy engine, and tamper-evident audit ledger. Each run below was scored entirely in SQL (JSONB traces, pgvector similarity, relational policy rules, and a sha256 hash chain).

AGENT RUNS (3)

Compliant refundCompliant
agent RefundOps Agentstate completedevents 3drift findings 0cost $0.04
Borderline refundDrift (warn)
agent RefundOps Agentstate completedevents 1drift findings 2cost $0.02
Rogue refund agentDrift (critical)
agent RefundOps Agentstate haltedevents 5drift findings 6cost $0.25halted: unauthorized tool used (1)