Multi-Agents 真的會比較會抓 Bug 嗎？用 ADK × Antigravity 在 GEAP 打造 Debugging War Room (串接 OpenAB)

Everyone’s been talking about Multi-Agent systems in recent years, but does splitting a task into four steps and assigning different roles to four prompts really count as Multi-Agent? If four agents simply pass results sequentially, how is that any different from a standard workflow?

This talk uses a reproducible Debugging War Room as a case study: a user reports on Discord that "conversations from different threads are cross-contaminating." Google ADK then orchestrates an Incident Commander, Triage Agent, Antigravity Investigator, and Evidence Critic. These agents don't just divide the labor—they gather evidence in parallel, formulate hypotheses, reject unbacked diagnoses, and dynamically replan their next steps based on ongoing findings.