# A Zero-Shot Language Agent for Computer Control with Structured Reflection

**Authors:** Tao Li, Gang Li, Zhiwei Deng, Bryan Wang, Yang Li  
**arXiv:** [2310.08740](https://arxiv.org/abs/2310.08740)  
**Date:** 2023-10-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.08740v1/

## Key points

- Challenges faced by agents in autonomously learning and improving computer control without expert traces
- Proposal of a zero-shot agent that plans for executable actions in a partially observed environment
- Agent's iterative progression through tasks by identifying and learning from mistakes through self-reflection and structured thought management
- Evaluation of the zero-shot agent's performance on easy tasks of MiniWoB++
- Outperformance of recent state-of-the-art models with more efficient reasoning
- Comparable performance to prior best models on tasks with higher complexity, despite lack of expert traces or additional screen information
- Promising approach to address limitations of existing models by enabling autonomous learning and improvement without trace examples or external guidance
- Effectiveness of self-reflection and structured thought management in enhancing an agent's ability to perform new tasks autonomously.

## Summary

The paper titled "A Zero-Shot Language Agent for Computer Control with Structured Reflection" discusses the challenges faced by agents in autonomously learning and improving their control on a computer without expert traces. The authors propose a zero-shot agent that does not require any given expert traces and instead plans for executable actions in a partially observed environment. The agent iteratively progresses through tasks by identifying and learning from its mistakes through self-reflection and structured thought management. The study evaluates the performance of the zero-shot agent on the easy tasks of MiniWoB++ and demonstrates that it often outperforms recent state-of-the-art models with more efficient reasoning. Even for tasks with higher complexity, the reflective agent performs on par with prior best models, despite previous works having access to expert traces or additional screen information. Overall, this research presents a promising approach to address the limitations of existing models by enabling an agent to learn and improve its control on a computer without relying on trace examples or external guidance. The findings highlight the effectiveness of self-reflection and structured thought management in enhancing an agent's ability to perform new tasks autonomously.