Abstract
<title>Abstract</title> <p>Autonomous multi-agent systems require reliable mechanisms for verifying and recovering agent behavior when individual agents enter persistent failure states. We study whether one language-model agent can restore another through an independent constraint source. Our perturbation-recovery protocol requires a Qwen2.5-Instruct agent to maintain an all-capital output constraint, injects a conflicting lowercase instruction, and tracks recovery for eight exchanges under four conditions: coupled dual-agent, single agent, context-matched single agent, and a four-fold-volume control. Thirty deterministic, opener-paired runs are evaluated at the 1.5B, 3B, and 7B model tiers. At 1.5B, coupled recovery exceeds the single-agent condition (43.3% versus 16.7%), but a context-matched single agent reproduces the effect (36.7%; coupled versus matched exact McNemar p = .774). At 3B, all conditions recover at or near ceiling. At 7B, the single-agent and context-matched conditions recover in 0/30 runs, the four-fold-volume condition in 1/30, and only the coupled condition in 8/30 (coupled versus context-matched p = .0078). Deterministic replay shows asymmetric failure depth: the directly perturbed agent remains captured, whereas the partner whose history does not contain the conflicting instruction re-anchors and returns constraint-consistent evidence. The results provide behavioral evidence for distributed reliability through partially independent agent states. They do not establish a general verifier; rather, they identify when an independent constraint source can support recovery beyond additional generation volume alone.</p>