Agent swarm experiment: prompt name order and model choice decide who leads
bibryam · x · 2026-11-10
Raindrop's 17-minute deep dive ran a simulated online store outage drill with 8 agents from four models (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.1 Pro, Kimi K3), all peers with equal permissions and no designated leader.
Key findings
- Many agents issued instructions, but only a few actually changed other agents' behavior and emerged as "leaders"
- Two factors determined leadership: the agent's underlying model, and the order its name appeared in the prompt — names were rotated across models to rule out bias
Takeaway: seemingly trivial configuration details like prompt name ordering significantly shape multi-agent coordination, with direct implications for swarm design.
More from coding & agent
- Ant Group's SkinMind: agentic RAG with graph probing for dermatology AI — antgroup · 2026-11-10
- Anthropic, OpenAI, Google, Vercel open-source their own AI skills on GitHub, barely used — blaizedsouza · 2026-11-10
- Ling-3.2-flash Adds Per-Request Tool Control: auto, any, tool, none Modes — Prudent-Food-2971 · 2026-11-10
- Lessons from shipping a remote MCP server with per-user keys — Salt_Bike898 · 2026-11-10
- Agent given a vocal-synth repo builds branching acoustic-tube singing synthesis — TinfoilTricorn · 2026-11-10
- A $0.63/hour AI employee in Slack: agent builds pages in 2 minutes, no briefs or handoffs — rubenhassid · 2026-11-10