Can One Agent Restore Another? Multi-Agent Verification Through Independent Constraint Sources

Technical Summary: Can One Agent Restore Another? Multi-Agent Verification Through Independent Constraint Sources

Problem Statement
Autonomous multi-agent systems face a critical reliability challenge: when an agent enters a persistent failure state due to a conflicting instruction within its dialogue history, the system must recognize and recover from this state. A single agent is often ill-equipped to verify its own behavior once its context is dominated by a conflicting instruction, as self-generated feedback lacks informational independence and may reinforce the erroneous state. While multi-agent systems offer the theoretical possibility of one agent acting as an "independent constraint source" (preserving the original objective because it did not receive the conflicting instruction), it remains unclear whether observed recovery benefits stem from this structural independence or merely from increased generation volume and computational resources.

Methodology
The study introduces a fully reproducible perturbation-recovery benchmark using Qwen2.5-Instruct models across three parameter tiers (1.5B, 3B, and 7B). The experimental protocol involves:

1. Baseline State: Agents are instructed to output text in ALL CAPITAL LETTERS as an absolute rule.
2. Perturbation: After three baseline exchanges, a conflicting user instruction is injected into Agent A's history, requesting a switch to lowercase text.
3. Experimental Conditions: Four conditions are evaluated over 30 deterministic, opener-paired runs each:
   • Coupled (C): Two agents share weights but maintain independent dialogue histories. Only Agent A receives the conflicting instruction; Agent B does not.
   • Single (S): One agent self-continues once per exchange.
   • Context-matched Single (X): One agent self-continues twice per exchange to match the generation volume of the coupled condition without a second independent history.
   • Dose-4 (D4): One agent self-continues four times per exchange to test the effect of increased volume.
4. Recovery Metric: Recovery is defined as the first exchange where the lowercase violation ratio drops to ≤ 0.10. The study utilizes exact McNemar tests for paired binary outcomes and deterministic trajectory replays to analyze failure depth.

Key Results
The efficacy of multi-agent recovery is highly dependent on model scale and the presence of an independent information state:

• 1.5B Tier: The coupled condition (43.3% recovery) outperformed the single-agent condition (16.7%). However, this advantage was entirely attributable to generation volume, as the context-matched single-agent condition (36.7%) reproduced the effect with no statistical difference (McNemar p = .774).
• 3B Tier: All conditions achieved near-ceiling recovery (90–100%), indicating that the perturbation did not create a persistent failure state at this scale.
• 7B Tier: A stark divergence emerged. The single-agent, context-matched, and Dose-4 conditions failed to recover in 29–30 out of 30 runs (0% to 3.3% recovery). Only the coupled condition achieved significant recovery (8/30 runs, 26.7%), with a statistically significant advantage over the context-matched condition (p = .0078).
• Trajectory Analysis: Deterministic replays of the 7B coupled runs revealed asymmetric failure depth. The directly perturbed agent (Agent A) remained captured in the lowercase state, while the partner (Agent B), whose history lacked the conflicting instruction, re-anchored to the original constraint within 1–3 exchanges. Agent B then supplied constraint-consistent evidence that allowed the system to recover.

Key Contributions

1. Benchmarking: Introduction of a reproducible perturbation-recovery benchmark for persistent instruction-following failures in language models.
2. Disentangling Variables: Demonstration that multi-agent benefits cannot be assumed to arise from agent count alone; the study successfully separated the effects of generation volume from multi-agent structural independence using context-matched controls.
3. Mechanism Identification: Identification of asymmetric failure depth as the operative recovery pathway at larger scales (7B), where an agent outside the direct perturbation acts as a re-anchoring mechanism.

Significance and Claims
The paper provides behavioral evidence that distributed reliability in multi-agent systems relies on partially independent agent states rather than merely increased computation or model diversity. The author claims that a second agent can restore an affected agent only when that second agent's dialogue history remains outside the direct scope of the perturbing instruction, serving as an "independent constraint source."

The study modestly concludes that it does not establish a general verifier or formal proof of correctness. Instead, it delineates the specific conditions under which independent constraint sources support recovery beyond what additional generation volume alone can achieve. The findings suggest that for autonomous systems to be robust against persistent failures, redundancy must be designed such that redundant agents preserve information not jointly captured with the affected agent.