Reading the “Generative Agents: Interactive Simulacra of Human Behavior” paper by Joon Sung Park, a PhD student in computer science at Stanford University, inspired me to conduct a more in-depth analysis and study on “AI Agent.” Recently, Google unveiled an AI agent based on Gemini 2.0, and Microsoft developed Magentic-One&Co-pilot Vision. As seen from examples from two big tech companies, we can see how it is already evolving from generative AI to ‘AI agents.’
<Concept of AI Agent>
To put it in the most straightforward form, ‘AI agent’ is a revolutionized form of generative AI. AI agents are AI-based applications that can recognize the environment, make decisions, and take action to achieve specific goals. The key is ‘agent,” that is, the ability of software to act independently. It can imitate human actions and autonomously carry out various tasks. This technology has four symbolic traits: autonomy, reactivity, proactivity, and social ability. While existing Generative AI tools have focused on content creation, such as text, images, and music, AI agents focus more on problem-solving and executing complex tasks.
According to the Deloitte report, 25% of companies using Generative AI are expected to start pilot projects or proof of concept for AI agents in 2025. By 2027, this rate is expected to increase by half. “Some AI agent applications may be introduced into real-world workflows from late 2025,” Deloitte said. “AI agents can increase the productivity of knowledge workers and make various workflows more efficient.
<How does it work?>
1. Data acquisition: The AI Agent acquires data through sensors, APIs, networks, and so on and preprocesses the data. This includes organizing data, extracting features, and standardizing tasks for subsequent analysis and processing.
2. Model training & Optimization: AI Agents use machine learning and deep learning techniques to train and optimize models based on large amounts of training data. AI Agents can learn the association between input data and output results through model training and extract features and patterns.
3. Decision Making: Learned and optimized models can be used for inference and decision-making to make corresponding decisions and actions based on input data and context. AI Agents can analyze and judge new inputs based on previously learned knowledge and patterns.
4. Feedback & learning: AI agents can continuously obtain and utilize feedback information for improvement and learning. This includes strategies such as evaluating output based on correct answers, adjusting the model’s parameters, updating training data, and more to improve the AI agent’s performance and efficiency.
5. Interaction & application: AI agents interact with users or other systems, receive input, and provide output. The application can cover several fields, such as intelligent secretaries, autonomous driving, financial services, and so on, and can be customized and applied according to specific tasks and scenarios.
<Real Life Examples>
JP Morgan: JP Morgan reduced fraud by 70% and saved $150 million annually by implementing AI agent’s fraud detection systems.
Microsoft: announced in November 2024 that it was adding automated agents to the M365 Copilot.
Google: Google’s AI agent achieved 85.4% sensitivity in diagnosing skin cancer, surpassing the accuracy of dermatologists.
<Risks>
- Energy consumption: Deloitte predicts that global data center electricity consumption could roughly double to 1,065 TWh by 2030, accounting for 4% of total global energy consumption. This necessitates investments in energy-efficient technologies and sustainable practices.
- Ethical and security concerns: Most AI agent tools claim to be safe and reliable, but their safety and reliability depend on the source of information used. Agents’ data sources can range from limited corporate data to open internet. The latter can affect Generative AI output and lead to errors or Hallucinations.
<References>
“Autonomous Generative AI Agents: Under Development.” Deloitte Insights, Deloitte, 18 Nov. 2024, www2.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2025/autonomous-generative-ai-agents-still-under-development.html.
Krakowski, Isabelle, et al. “Human-AI Interaction in Skin Cancer Diagnosis: A Systematic Review and Meta-Analysis.” Npj Digital Medicine, vol. 7, no. 1, 9 Apr. 2024, https://doi.org/10.1038/s41746-024-01031-w.
“Deloitte Global’s 2025 Predictions Report: Generative AI: Paving the Way for a Transformative Future in Technology, Media, and Telecommunications.” Deloitte, 19 Nov. 2024, www.deloitte.com/global/en/about/press-room/deloitte-globals-2025-predictions-report.html.
Park, Joon Sung, et al. “Generative Agents: Interactive Simulacra of Human Behavior.” ArXiv:2304.03442 [Cs], 6 Apr. 2023, arxiv.org/abs/2304.03442.
Leave a Reply