Demonstrating Autonomous Purchase Automation with Qwen3.8 27B on Local Hardware
In an impressive showcase of AI and automation capabilities, a fully local deployment of the language model Qwen3.8 27B successfully executed an entire online shopping task—from search to purchase—without relying on cloud services.
Scenario Overview
Faced with a simple yet practical challenge—restocking A4 paper—the user tasked the local AI agent with identifying affordable options on Amazon and completing the purchase independently. The goal was to assess how well a self-contained AI system could handle web navigation, decision-making, and transaction execution.
Technical Setup
The system configuration comprised:
- Qwen3.8 27B: A powerful local language model running entirely on the user’s hardware.
- TensorSharp: An inference engine enabling local large language model computations.
- Agent Runtime: Facilitating autonomous task management and decision-making.
- Playwright Integration: Automating browser interactions based on generated instructions.
This architecture ensured that all reasoning, inference, state management, code generation, and execution of tools occurred locally, with only the interaction with Amazon’s website performed through the web browser.
Execution and Performance
The autonomous agent initiated a web search, navigated Amazon’s product listings, compared prices, and added the chosen items to the cart. Throughout the process, the system required only two manual interventions: logging into Amazon and confirming the final purchase.
The entire task was completed in approximately 24 minutes, with the detailed process captured in a speeded-up video at 8× real-time. The clip reveals the sequence of browser commands and the Playwright-generated code, illustrating the step-by-step decision-making process undertaken by the AI.
Noteworthy Aspects
- Privacy and Data Privacy: The entire inference and reasoning processes remained on local hardware, enhancing data security.
- Flexibility and Control: The setup allows for modifications and extensions without relying on cloud APIs.
- Current Limitations: While the desktop implementation performed efficiently, the iOS version (TensorAgent) still faces challenges related to inference speed and browser restrictions, indicating areas for future improvement.
Conclusion
This demonstration underscores the potential of fully local AI systems to automate complex tasks traditionally requiring cloud-based services. As hardware capabilities advance, such autonomous agents could become standard tools for managing routine online activities securely and efficiently.
Additional Resources
For those interested in exploring or replicating this setup, the project code is available on GitHub: https://github.com/zhongkaifu/TensorSharp.
Harnessing the power of local inference and automation, this example marks a significant step toward privacy-preserving, self-reliant AI solutions for everyday tasks.
