AUDIT TRAILLegal asks what the agent did and why. The answer lives in four tools.
Every detection and every healing action is signed and replayable. One query returns the chain of evidence behind any model call — not stitched together from Datadog plus Langfuse plus Sentry plus screenshots.
●signed remediation · replayable · 1-query trail
SCOPE CREEPThe agent acquires more agency than it was designed for. Nobody notices.
Pisama baselines permissions and tool grants against the deploy snapshot. The moment the running agent reaches for capability it was not shipped with, the specification-compliance detector trips.
●specification_compliance · F1 0.966 · per-turn
SYCOPHANCYThe agent tells customers what they want to hear instead of what is true.
Agreement-shaped responses that contradict the underlying evidence get flagged at the turn they happen. Brand damage caught before it leaves the agent — not after a Twitter thread lands.
●sycophancy · F1 0.902 · per-turn
CASCADING FAILUREOne agent’s bad output poisons every downstream agent in the graph.
Graph-aware circuit breakers cap blast radius at the agent boundary. The consensus-collapse detector (F1 0.967) catches voting ensembles converging on a wrong answer before the answer propagates.
●consensus_collapse · F1 0.967 · graph-aware breakers