DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents

Researchers have identified a failure mode they call "loops" in long-horizon language-model agents, where an agent re-issues a tool call it already made and receives the same observation, adding no new information. The researchers report that loops appear natively in Qwen, Llama, and Gemini agents, and propose DeLoop, which uses repeated actions as self-supervision for long-horizon agents.

Language-model agents that run long tasks often redo work they have already done: they re-issue a tool call they made earlier and get back exactly the same observation, a step that adds no new information. We identify this failure as a loop. Loops appear natively in Qwen, Llama, and Gemini agents an