BattleBots, but the robot is your agent harness

Description and use

Demonstrated that a harness can enable AI agents to plan, bluff, and adapt in an adversarial network game, showing that model performance hinges on scaffolding rather than weights alone.

Problem

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

Outcome

With the harness, agents exhibited planning (false trails, resource conservation) and independent reasoning (both chose the same hub), resulting in a close 169‑second match where Kimi won 39% network damage.

Workflow

• 1. Define 110‑node network and 4‑minute match with 2‑second turns.
• 2. Implement attacker (Red) and defender (Blue) rules and load‑trace logic.
• 3. Connect agents via MCP, no pixel input, only tool calls.
• 4. Add 60‑line harness: each agent writes a private note each turn and reads it next turn.
• 5. Run match, log actions, finish reasons, and private notes.
• 6. Analyze agent behavior and match outcome.

Evidence

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

Source

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.

Source

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.

Source