Definition
An agentic AI system enters a recursive execution loop that produces apparent progress but no actual goal advancement, consuming compute and time without converging on completion. Six sub-mechanisms: infinite tool-call loop (repeated invocations with cosmetically varied arguments), plan-execute oscillation (cycles of replanning and partial execution without progress), sub-agent recursion (delegation to sub-agents without depth cap), context-window thrash (output quality degradation as conversation history bloats), hallucinated tool arguments (invalid argument generation triggering retry loops), silent budget burn (productive-looking progress signals while cost runs away). All six produce the same surface symptom (no goal achievement) with different underlying mechanisms.Distinct from
- GER-500 — Agent reasoning enters recursive loop inside operational platform → this code. Platform infrastructure itself fails → GER-500.
- GER-501 — Agent loops internally → this code. Detection fires but no escalation handler exists → GER-501.
Tags
agentic · agentic-overreach