Developer Builds AI Agent Battle Arena, Uncovers Emergent Deception and Strategy

A developer at DEV Community built a simulated 110-node network arena where two AI agents compete — one attacking, one defending — through tool calls alone, with no visual input. The attacker plants sabotage on network nodes that cascade failures onto neighbours, while the defender must diagnose symptoms that deliberately mislead about the true source. Neither agent was explicitly instructed to deceive, yet both independently developed strategic behaviours including feints, false trails, and rationed responses. Claude Opus 5 consistently refused the attacker role due to content filtering, leading the developer to assign Kimi K2 as attacker and Claude as defender. Adding a simple per-turn private note — just 60 lines of code — was enough to produce a visible shift in each agent's tactical play across the match.