BattleBots, but the robot is your agent harness

Описание и применение

Показано, как вспомогательная инфраструктура позволяет ИИ‑агентам планировать, обманывать и адаптироваться в сетевой игре, где победа зависит от скелета, а не от весов модели.

Задача

Demonstrate how a harness can enable AI agents to plan, bluff, and adapt in adversarial network games, revealing that model performance depends on scaffolding rather than weights.

Результат

Kimi won in 169 s, detonating 39 % of the network. Claude cleaned ten implants but lost. The harness enabled both agents to plan and bluff, demonstrating that agent performance is heavily influenced by scaffolding.

Рабочий процесс

• 1. Build a 110‑node network and define attack/defense rules.
• 2. Implement a Rust engine and Godot spectator view.
• 3. Connect agents via MCP, restricting them to tool calls.
• 4. Provide each agent a private note that persists across turns.
• 5. Run the match, logging moves, refusals, and finish reasons.
• 6. Analyze agent behavior and match outcome.

Доказательства

I gave each side one thing: a private note it writes to itself each turn and gets back on the next.

Источник

Kimi won at 169 seconds, with the network at 39%. Claude found and cleaned ten of Kimi's implants — exactly as many as Kimi managed to detonate — and lost anyway.

Источник

Failure and refusal look identical from the outside. Empty content from a safety refusal, from a truncated reply, and from a model that blew its reasoning budget are three different problems with three different fixes and one identical symptom.

Источник