Search the Atlas

Search risks, controls, and glossary terms

MediumAgenticExcessive Action HorizonDemonstrated

Runaway Multi-Step Planning

Agent Autonomy

Description

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 evidenceDemonstrated

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

Related risks in Agent Autonomy