Multi-Agent Arbitration for Conflicting Code Diffs

Multi-Agent Arbitration for Conflicting Code Diffs

When two agents propose incompatible diffs, blindly picking the latest output is risky. Arbitration needs structured scoring and conflict-aware merge strategy.

Step 1: normalize agent proposals into comparable artifacts

{
  "agent": "planner",
  "files_changed": 4,
  "tests_passed": 27,
  "risk_flags": ["migration"]
}

Step 2: apply weighted scoring rules

score = 0.5 * tests + 0.2 * lint + 0.2 * safety - 0.1 * risk_flags

Step 3: route close-score conflicts to human review

if abs(score_a - score_b) < 0.05:
    return "manual_review"

Pitfall

Optimizing for speed only. Fast merges with unresolved semantic conflicts create expensive regressions.