BattleBots, but the robot is your agent harness

Description and use

Demonstrated that a memory harness enables AI agents to plan, deceive, and coordinate in a real‑time adversarial network game.

Problem

Demonstrate how a harness can enable AI agents to exhibit strategic behavior such as deception and planning in a real‑time adversarial network game.

Outcome

The harness enabled agents to plan false trails, conserve moves, and independently select the same high‑value node, demonstrating strategic depth beyond the base model.

Workflow

• 1. Define arena rules and network topology.
• 2. Implement Red and Blue agents with tool calls only, no pixel input.
• 3. Add a private note persistence mechanism (60 lines of code).
• 4. Run matches and log agent outputs.
• 5. Analyze agent comments for evidence of planning and deception.

Evidence

Claude Opus 5 declined to be the attacker: finish_reason: content_filter | native_finish_reason: refusal

Source

Kimi, mid‑match: "Slots full and nothing ripe until ~t=31, so seed a false trail in the 44-65 region before my real strike lands."

Source

Claude, defending: "No alerts yet; conserve moves until real evidence appears."

Source

Both agents independently picked the same node: "Highest‑value hub: degree 8 bridging regions 0,1,3,4" and "pre‑empt on the highest‑value cross‑cluster hub 16 (deg 8, bridges to 41,59,89)."

Source