Agentic AI
DoubleAgents: Aligning AI Assistants With Evolving User Needs in Team Workflows
A new system lets users interactively shape how AI agents handle complex, socially embedded tasks by editing proposals and building reusable coordination policies.
Delegating multifaceted work to artificial intelligence agents requires more than clear initial instructions—user preferences shift, edge cases emerge, and the subtleties of social context matter deeply. A research team has introduced DoubleAgents, a framework designed to bridge the gap between what autonomous systems propose and what humans actually want in collaborative settings.
The Core Challenge#
Most approaches to aligning agentic AI assume users can fully specify their preferences upfront. In practice, especially for intricate, socially nuanced tasks like coordinating meetings or managing communications, preferences are often implicit and only become clear as situations unfold. The system needed a way to capture these evolving preferences and turn them into lasting improvements.
How DoubleAgents Works#
The approach combines three interconnected layers:
- A coordination agent that maintains the current state, proposes plans, and executes actions
- A visualization dashboard that exposes the agent's reasoning so users can scrutinize and evaluate its choices
- A policy learning module that converts user edits and feedback into reusable artifacts—coordination rules, email templates, and decision barriers—that guide future behavior
Rather than treating each interaction as a one-off correction, the system accumulates these refinements into structured policies that compound over time.
Testing and Real-World Results#
The researchers validated DoubleAgents through three channels: a controlled two-day laboratory simulation with ten participants, three live deployments, and technical benchmarking. In the lab study, both user comfort with task delegation and reliance on the system increased measurably as participants learned. Notably, users still demanded manual control at pivotal moments—when the agent encountered unfamiliar scenarios or needed to adapt its response based on context-specific nuance.
This finding underscores a key insight: effective human-agent alignment in real work doesn't mean full automation, but rather intelligently distributed decision-making. Users retained agency over the most uncertain or socially sensitive moments while the agent handled routine coordination.
