TLDR: DoubleAgents is an AI-powered planning tool designed to build human trust in proactive AI for complex human coordination tasks like event scheduling. It uses a dual-agent system: a Coordination Agent for planning and execution based on user policies, and Simulated Respondents that mimic realistic human behavior for safe testing. Key features include transparency, user control, adaptive learning from user edits, and edge case detection. Studies show users gain trust and comfort with AI delegation through consistent performance, control, and clear explanations, making it a valuable tool for real-world coordination.
The rise of AI agents promises to make our lives more efficient by automating complex tasks. However, a significant hurdle remains: building human trust in systems that act proactively on our behalf, especially for delicate human coordination challenges like organizing events or scheduling speakers. A new research paper introduces “DoubleAgents,” an innovative agentic planning tool designed to foster this crucial trust through transparency, control, and adaptive learning.
Authored by Tao Long, Xuanming Zhang, Sitong Wang, Zhou Yu, and Lydia B. Chilton from Columbia University, the paper delves into how AI can assist with intricate human coordination tasks, which often involve unpredictable responses, social nuances, and continuous replanning. Traditional scheduling tools fall short because they lack the ability to handle ongoing negotiation and relationship maintenance.
Understanding DoubleAgents: A Dual-Agent Approach
DoubleAgents tackles this challenge by integrating two complementary types of generative AI agents. First, a Coordination Agent uses a “Reasoning and Acting” (ReAct) loop to plan and execute actions, such as drafting and sending emails to potential speakers. This agent operates based on user-defined policies and can call various tools to gather context and perform actions.
Second, Simulated Respondents are LLM-based agents that mimic realistic human behavior. They respond to emails based on predefined personas (e.g., a busy professor, a slow responder, a speaker with specific preferences), generating varied and lifelike scenarios. This simulation environment allows users to safely test and refine the system’s policies and calibrate their reliance before live deployment.
Key Mechanisms for Building Trust
The system incorporates several features to build trust:
- Transparency and Control: Users can observe and intervene at every step of the planning and execution process. The system provides rich state visualizations, including an assignment tracker, a communication history viewer, and a calendar view, offering a clear overview of progress.
- Policy-Driven Planning: The Coordination Agent follows explicit, human-articulated policies that guide its behavior. Users can review, modify, or add policies, ensuring the AI’s actions align with their values and preferences.
- Implicit User Preference Learning: DoubleAgents learns from user edits to plans, actions, and emails. If a user refines an email’s tone or removes a suggested action, the system incorporates this feedback into future generations, adapting to the user’s communication style and preferences over time.
- Edge Case Detection: The system is designed to identify situations that fall outside existing policies—such as a speaker requesting a virtual presentation option when only in-person was planned. In such “edge cases,” DoubleAgents flags the issue and escalates it to the user for clarification and guidance, preventing autonomous mishandling of delicate situations.
Also Read:
- Simulating Intelligent Agents: A New Platform for Event Coordination and Realistic Interactions
- HARMONIC: A New Robotic Architecture for Intelligent Human-Robot Collaboration
Evaluation and Real-World Relevance
The researchers conducted a two-day lab study with 10 participants and a week-long deployment with two real-world event organizers. The results were compelling: participants initially hesitated to delegate tasks to the AI but grew more comfortable and reliant as they experienced the system’s transparency, control, and adaptive learning during simulated cases. The simulation rounds were particularly highlighted as crucial for building confidence, allowing users to anticipate and address potential real-life problems.
In live deployments, organizers found DoubleAgents highly relevant for their email-heavy coordination tasks. They appreciated the system’s ability to generate personalized emails and adapt to their writing styles. The effort required for setup scaled appropriately with task complexity, and the system’s consistent performance, combined with visible policies and edge-case flagging, fostered trust. One organizer noted that understanding the system’s capabilities and limitations over time was key to their comfort.
The paper concludes by distilling three core trust-by-design mechanisms: consistency (reliable, policy-driven behavior), control (user intervention and adaptive learning), and understanding (transparency through visualizations and explanations). These mechanisms align with how humans build trust with each other and address common challenges in human-AI interaction design.
For a deeper dive into the technical details and findings, you can read the full research paper: DoubleAgents: Exploring Mechanisms of Building Trust with Proactive AI.
