What are LLM-based Autonomous Agents

Autonomous agents represent self-operating systems that exhibit varying degrees of independence. Recent research highlights the remarkable capacity of LLMs to imitate human intelligence, a feat achieved through the combination of extensive training datasets and a substantial array of model parameters. This research article provides a comprehensive study of the architectural aspects, construction techniques, evaluation methods, and challenges associated with autonomous agents utilizing LLMs.


LLMs have been utilized as core orchestrators in the creation of autonomous agents, aiming to replicate human decision-making processes and enhance artificial intelligence systems. The above image constitutes an illustration of the growth trend in the field of LLM-based autonomous agents. It is interesting to note how the X-axis switches from years to months after the third point. Essentially, these LLM-based agents are evolving from passive language systems into active, goal-oriented agents with reasoning capabilities.

LLM-based Autonomous Agent Construction

In order to demonstrate human-like capabilities effectively, there exist two significant aspects to note:

  1. Architectural Design: Selecting the most suitable architecture is important for harnessing the capabilities of LLMs optimally. Existing research has been systematically synthesized, leading to the development of a comprehensive and unified framework.
  2. Learning Parameter Optimization: To enhance the architecture’s performance, three widely employed strategies have emerged:
  • Learning from Examples: This approach involves fine-tuning the model using carefully curated datasets.
  • Learning from Environment Feedback: Real-time interactions and observations are leveraged to improve the model’s abilities.
  • Learning from Human Feedback: Human expertise and intervention are capitalized upon to refine the model’s responses.

LLM-based Autonomous Agent Application

The application of LLM-based autonomous agents across various fields signifies a fundamental shift in how we address problem-solving, decision-making, and innovation. These agents possess language comprehension, reasoning, and adaptability, leading to a profound impact by providing unmatched insights, support, and solutions. This section largely delves into the transformative effects of LLM-based autonomous agents in three distinct domains: social science, natural science, and engineering.

LLM-based Autonomous Agent Evaluation

To assess the effectiveness of the LLM-based autonomous agents, two evaluation strategies have been introduced: subjective and objective evaluation.

  • Subjective Evaluation: Some potential properties, like agent?s intelligence and user-friendliness, cannot be measured by quantitative metrics as well. Therefore, subjective evaluation is indispensable for current research.
  • Objective Evaluation: Utilising objective evaluation presents numerous advantages in comparison to human assessments. Quantitative metrics facilitate straightforward comparisons among various approaches and the monitoring of advancements over time. The feasibility of conducting extensive automated testing enables the evaluation of numerous tasks instead of just a few.

Finally, although previous work has shown many promising directions, this field is still at its initial stage, and many challenges exist on its development road, including role-playing capability, Generalised Human Alignment, Prompt Robustness etc. In conclusion, this survey provides us with a detailed study of everything that is in the know about LLMs-based Autonomous agents and provides us with a systematic summary of the same.