Non-Deterministic Agent Behaviour
Accountability & GovernanceDescription
Agent outputs are non-deterministic even for identical inputs due to stochastic planning, memory retrieval, tool responses, and external state. Prevents reliable testing, auditability, and regulatory compliance verification.
Regulatory audit cannot reproduce the agent's decision because execution state and tool responses were not captured.
Multi-agent orchestrator-subagent architectures are nascent; no published research or confirmed incident documents a subagent successfully exploiting orchestrator trust to escalate privileges in a production deployment. The risk is architecturally plausible but has not been empirically demonstrated in a controlled experiment targeting this specific trust boundary.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Deterministic execution modes for audit
- comprehensive state logging
- environment snapshotting
- reproducible agent execution frameworks.
Detection signals
Reproducibility Score; behavioural variance across identical runs; audit trail completeness.