SANZARU: three monkeys cooperating to defuse a bomb

SANZARU / A BENCHMARK FOR AI TEAMWORK

Can three AI agents defuse a bomb together
with limited perception?

🙉 Deaf sees the bomb🙊 Mute reads the manual🙈 Blind cuts the wires

Yiping Li · Zhiyu An · Wan Du UC Merced

01 / Two asymmetries

No one can defuse it alone.

TASK / BASE-1#2
DISPLAY 0LIGHT green
01yellow
02green
03red

02 / An example

One bomb, start to finish.

Gemini-3.8-Flash · Thinking ON

base-1#2 · run-1 · 3 rounds

Defused
Three perspectives

All recorded messages and actions, in turn order.

Loading conversation…

03 / Results

How well do models cooperate?

Team success

216 episodes / model

What if we remove the communication asymmetry?

Qwen3.8-27B · Thinking ON

With communication asymmetry62.5%
Without communication asymmetry82.4%

04 / Failure modes

Why do the bombs explode?

Centre numbers show the number of failed episodes.

Gemini-3.8-Flash

Thinking ON

Gemini: 30 losses. Communication 30%, reasoning 60%, execution 10%.

GLM-5.3-Flash

Thinking ON

GLM: 81 losses. Communication 38%, reasoning 23%, execution 38%.

GPT-5.6-Terra

Thinking ON

Terra: 97 losses. Communication 46%, reasoning 52%, execution 2%.

DeepSeek-V4.1-Flash

Thinking OFF

DeepSeek: 110 losses. Communication 54%, reasoning 29%, execution 17%.

GPT-5.6-Luna

Thinking ON

Luna: 128 losses. Communication 62%, reasoning 27%, execution 11%.

Claude-Sonnet-5

Thinking OFF

Claude: 153 losses. Communication 75%, reasoning 14%, execution 10%.

Qwen-3.8-Flash

Thinking OFF

Qwen: 172 losses. Communication 73%, reasoning 13%, execution 14%.

All models

Pooled losses

All models: 771 losses. Communication 60%, reasoning 26%, execution 14%.

Communication

  • No shared code
  • Misalignment
  • Omission
  • Slow protocol

Reasoning

  • Wrong answer
  • Wrong round

Execution

  • Hesitation
  • Impatience

05 / Perspective taking

Can agents take their teammates’ perspectives?

The prompts explicitly state who can see, hear, speak, and act. Agents still make plans and send messages that ignore these limits.

67%

of failed episodes 516 / 771

Assuming abilities a teammate does not have.

Naming a colour to Blind, or expecting Mute to speak.

19%

of failed episodes 148 / 771

Sending through a channel a teammate cannot receive.

Speaking to Deaf, or gesturing to Blind.

06 / Real episodes

Inside the conversation.

Gemini-3.8-Flash · Thinking ON

base-3#3 · run-3 · 4 rounds

Defused
Three perspectives

All recorded messages and actions, in turn order.

Loading conversation…