BattleBots, but the robot is your agent harness

Описание и применение

Показано, как добавление постоянной памяти и правильного скелета повышает стратегические возможности ИИ в сетевой игре с каскадными сбоями.

Задача

Demonstrating how harness design (memory, prompts, tool ergonomics) affects AI agent performance in adversarial network games, revealing that model behavior can be driven by scaffolding rather than inherent intelligence.

Результат

Adding persistent memory via a private note enabled AI agents to plan, deceive, and coordinate, transforming a simple cascading failure demo into a watchable adversarial game that highlighted harness design over model capability.

Рабочий процесс

• 1. Build a 110-node network arena with defined attack (plant, arm, detonate) and defense (probe, trace) actions.
• 2. Configure agents to communicate solely via tool calls, no visual input.
• 3. Initially provide agents only with system prompt and recent moves.
• 4. Observe agent behavior; note lack of planning and consistent refusal by Claude Opus 5 to attack.
• 5. Introduce a private note mechanism (60 lines of code) that persists across turns.
• 6. Re-run the match; agents now generate plans, mislead, and coordinate actions.
• 7. Analyze logs to confirm strategic behavior and shared node selection.
• 8. Document lessons on harness design, error handling, and agent memory.
• 9. Propose future enhancements (multiple attacks, costs, levels).

Доказательства

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

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