A developer announced on February 16, 2026, an AI automated development workflow claimed to run continuously for over 10 hours to complete complex full-stack software development tasks. The core of this solution combines Alibaba’s latest open-source model, Qwen-3.5-Plus, with a hierarchical agent system architecture.
Challenge and Architecture 1.0: Solving AI “Amnesia”
Current AI coding assistants commonly suffer from context length limitations, leading to logical inconsistencies or forgetting previous instructions during long, complex development tasks, which requires significant manual supervision. To address this, the developer designed an “AI Autonomous Development System.” Its core idea is a “worker rotation model” that manages state through an external file system rather than relying on the model’s internal memory.
The 1.0 version of this system consists of the following main files:
- feature_list.json: Acts as a structured task list, defining the features to be developed and their statuses.
- progress.txt: Records high-level development decisions and progress logs.
- CLAUDE.md: Serves as a system-level prompt, ensuring the AI restores its workflow and behavioral guidelines each time it starts.
- run_autonomy.py: An automated loop script responsible for reading tasks, invoking the AI model for execution, and updating the status upon task completion, enabling unattended, continuous development.
Core Engine: Practical Application of Qwen-3.5-Plus
The workflow selects Qwen-3.5-Plus as its core AI engine. Practice has shown that the model’s features are highly suitable for high-intensity Agent tasks. Firstly, its Mixture of Experts (MoE) architecture, with 397B total parameters but only activating about 17B for each inference, strikes a balance between high performance and efficiency. Secondly, as an open-source model, its deployment cost on platforms like Alibaba Cloud’s Bailian is far lower than closed-source models of comparable performance, making long-duration, high-frequency API calls economically viable. Lastly, the model excels in Agent capabilities such as following instructions and breaking down complex tasks, enabling it to accurately parse and execute tasks defined in feature_list.json.
System Upgrade 2.0: Introducing an Agent Team for Collaborative Development
When handling more complex production environment integration tasks, the single-threaded Agent model revealed efficiency bottlenecks. For example, backend API development and frontend UI adjustments could not be done in parallel, and the AI tended to get stuck in ineffective self-repair loops when encountering test errors. To overcome this, the developer introduced the “Agent Team” architecture, upgrading the single AI into a “AI development team” with a clear division of labor.
The team includes four different roles:
- Lead Agent (CTO): Responsible for overall task planning, decomposition, and Code Review; does not write code directly.
- @backend-integrator: Focuses on Python, FastAPI, and database interaction, responsible for implementing backend APIs.
- @frontend-polisher: Specializes in Next.js and UI components, responsible for frontend development and integration.
- @qa-engineer: Specializes in Playwright end-to-end testing. When a test fails, it does not attempt to fix the code but instead submits the error logs and report directly to the appropriate specialist Agent.
Through this architecture, the system successfully simulates the collaborative work model of a real development team. For instance, when integrating with a Sora2 video generation API, the backend Agent handles API polling logic while the frontend Agent concurrently develops a loading state UI. The QA Agent then automatically performs full-link verification after the feature is completed and accurately feeds back issues like API key errors to the backend Agent, ultimately achieving a highly efficient parallel development and problem-resolution cycle.
