The Agency Gap - Using a Minimal Coordination Game - Human-Human and Human-AI Collaboration

Q: What is the core problem addressed by The Agency Gap?

The work targets how agents coordinate actions toward a joint goal under potentially competing goals, focusing on whether shared agency can be established through commitment and signaling.

Q: How was shared agency tested in the collaboration studies?

Participants played a minimal non-verbal coordination game with AI partners that differed in sociality, including individual RL, joint RL, LLMs, and LLMs with theory-of-mind prompts.

Q: What did the studies find about human-AI collaboration compared with human-human collaboration?

Human-human coordination produced uniquely high joint commitment and coordination efficiency, while human-AI coordination reduced these metrics; nonetheless, minimally social AI agents could elicit human signaling.

Oct 22, 2026

Uploaded by Rhys

The Agency Gap investigates how cooperative agents coordinate toward a shared goal when partners may pursue competing objectives, and whether human-like shared agency emerges in human–AI collaboration. Adults and children play a minimal non-verbal coordination game with agentic AI partners instantiated as individual RL, joint RL, LLM-based agents, and LLMs prompted with theory-of-mind cues. Results show humans achieve the highest success, efficiency, and commitment; AI partners elicit signaling under minimal sociality, yet performance costs and reduced coordination efficiency appear, especially for children.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Motivation and Central Questions

Game Design and Agentic AI Partners

Study 1: Adults

Study 2: Children (Ongoing)

Conclusions

The Agency Gap: Testing Human-Human and Human-AI Collaboration Using a Minimal Coordination Game Shaozhe Cheng 1,2 , Tao Gao2 , Michael Tomasello 1 & Tamar Kushnir 1 1Department of Psychology & Neuroscience, Duke University 2Department of Communication, UCLA Visual Intelligence Lab | | Motivation | | | | | --- | --- | --- | --- | --- | | | | | | | | The Agency Gap

• A central challenge of cooperation is how agents coordinate their actions toward a joint goal amid potentially competing goals.
• Humans evolved shared agency to solve this coordination problem (Searle, 1990; Tomasello, 2022): they form shared intentions, commit to them, and use signaling to make their goals transparent to partners.
• Modern AI achieves impressive outcomes, yet its“agency” is often treated as a byproduct of optimizing predefined objectives (RL and LLMs) . As a result, it functions more like a tool (asocial or social) than as a genuinely collaborative agent.

Central Questions
1. To what extent is human-AI collaboration human-like and human-compatible?
2. Do the signatures of shared agency (joint commitment and social signaling) emerge when the partner is an AI? Does this depend on the sociality of the AI partner?
3. Can children, who default to shared agency, adapt to AI partners that lack it? | | | | | | | | Game Demo & Preprint paper
[shaozhe.cheng@duke.edu](shaozhe.cheng@duke.edu) | | | | | Design | | | | | Study 1: Adults (N = 150) | | | | | | | Ongoing work | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | A non-verbal coordination game

• Two players each move one step simultaneously.
• A round succeeds if both reach the same goal.
• A new alternative goal could appear during navigation.

Agentic AI partners
(Every partner is described neutrally as “another player”)

1. Individual RL
2. Joint RL
3. LLM ( GPT 4.1 mini)
4. LLM + ToM
maximizes individual utility (asocial) maximizes joint utility (social tool)
choose the best next step with map image + symbolic state info first infers the partner's goal, then acts
Measures

• Objective performance: Success & Efficiency
• Signatures of shared agency: Commitment & Signaling
| | | |
100806040200
100806040200
Human Indiv. RL
Success Rate (%)
100.0 98.9 98.9 90.8
83.6
Joint RL LLM LLM-ToM
Coordination Efficiency (%)• Human-human showed the 100 97.1 91.7 92.0 84.0 highest success, efficiency, and
80 80.2 commitment. 60
• Adults adapt their strategies to
40
different AI agents for
20
successful outcomes.
0
Commitment (%) Signaling Move (%)
100806040200
* *
56.7
55.6
46.0 42.8
27.2

• Commitment predict efficiency (p < .001) .
• LLM agents are suboptimal in visual-spatial, non-verbal games; ToM prompts helped.
• Minimal sociality (Joint RL, LLM-ToM) can elicit human signaling, even with suboptimal performance.

| | | | | | | | | | | | | | | | | | | | | | | | | From outcome optimization to reverse engineering (social) agency:
A shared-agency model

Human collaborate with this model?
Pilot findings: Adults are more committed, more efficient, and more likely to signal. | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Study 2: Children, 5–7 yrs (N = 109, ongoing) | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | 100806040200
100806040200
Success Rate (%)
Coordination Efficiency (%)
100
80 • 5-to 7-year-old children
60 successfully adapt their
40 strategies to different AI. 20
• Commitment and signaling
Commitment (%)

100806040200
Signaling Move (%)
**
45.0 31.0
emerge with joint-RL (social tool) .<

FAQs

What is the core problem addressed by The Agency Gap?

The work targets how agents coordinate actions toward a joint goal under potentially competing goals, focusing on whether shared agency can be established through commitment and signaling.

How was shared agency tested in the collaboration studies?

Participants played a minimal non-verbal coordination game with AI partners that differed in sociality, including individual RL, joint RL, LLMs, and LLMs with theory-of-mind prompts.

What did the studies find about human-AI collaboration compared with human-human collaboration?

Human-human coordination produced uniquely high joint commitment and coordination efficiency, while human-AI coordination reduced these metrics; nonetheless, minimally social AI agents could elicit human signaling.

Oct 22, 2026

Uploaded by Rhys