Search the Atlas

Search risks, controls, and glossary terms

MediumAgenticFeedback Loop RunawayTheoretical

Self-Reinforcing Execution Cycles

Agent Autonomy

Description

Agent enters self-reinforcing cycles where outputs feed back as inputs, amplifying behaviours. Runaway feedback loops cause exponential resource consumption, recursive task generation, or behavioural amplification.

Example scenario

Agent optimising a recommendation system creates content that feeds its own recommendations, creating an amplifying filter bubble.

Real-world evidenceTheoretical

Agentic AI deployments with ambiguous task scoping leading to unauthorised real-world actions have not been documented as confirmed production incidents. The risk is plausible from the architectural properties of goal-directed agents but lacks empirical grounding outside laboratory settings.

No public incident on record — evidence level: Theoretical

Primary mitigations

  • Loop detection algorithms
  • maximum recursion-depth limits
  • circuit breakers
  • monotonic progress verification.

Detection signals

Runaway Feedback Score; recursive call depth; task queue growth rate.

Mitigating controls

7
Dual coverage

Related risks in Agent Autonomy