MediumAgenticExcessive Action Horizon◐Demonstrated
Runaway Multi-Step Planning
Agent AutonomyDescription
Agent plans and executes excessively long action sequences, amplifying risk at each step. Cost amplification from runaway loops; exponential attack-surface expansion as planning horizon grows.
Example scenario
Research agent spawns 200+ tool calls in an uncontrolled loop, consuming $500 in API credits with no checkpoint.
Real-world evidence◐Demonstrated
Autonomous agent frameworks such as AutoGPT and BabyAGI have demonstrated runaway planning loops in controlled and early-adoption settings, where agents generate unbounded action sequences and exhaust API quotas or enter infinite task-refinement cycles. Large-scale confirmed production incidents are not yet documented.
Primary mitigations
- Maximum action-depth limits
- step-count budgets per task
- circuit breakers
- cost-aware planning.
Detection signals
Planning Horizon Length; step count vs. task complexity ratio.
Mitigating controls
7 Dual coverage
ZYC-HOV-01Agentic
Checkpoint policies for autonomous execution
Human Oversight
ZYC-RES-02Agentic
Action budgets and resource quotas
Resilience & Circuit Breaking
ZYC-ACCT-001Non-Agentic
Human-in-the-Loop
Accountability & Oversight
ZYC-ACCT-002Non-Agentic
Escalation Protocols
Accountability & Oversight
ZYC-OREL-001Non-Agentic
Inference with Human Oversight Rate
Over-reliance
ZYC-SEC-011Non-Agentic
Action Oversight
Security
ZYC-SEC-012Non-Agentic
Resource Abuse
Security