Technical Summary: FullStack-Agent

Problem Statement

Current Large Language Model (LLM)-powered code agents struggle to assist non-expert users in developing complex, production-level full-stack web applications. Existing systems predominantly generate frontend-only interfaces, masking the absence of real backend data processing and storage with superficial visual effects. While a form may appear to submit successfully, no data is actually processed or stored due to missing backend and database implementations. Furthermore, constructing production-grade full-stack applications presents three distinct challenges that current agents fail to address:

1. Complex Codebase Navigation: Real-world frameworks (e.g., Next.js, NestJS) involve large, intricate codebases requiring efficient navigation and precise localization of obscure bugs.
2. Expert Reasoning: Full-stack workflows demand long-term reasoning, skilled tool invocation, and deep mastery of constantly updating web packages and dependencies.
3. Inadequate Evaluation: Existing benchmarks (e.g., WebGen-Bench) rely on GUI-agent judges that primarily assess UI-level interactions, failing to detect missing or incorrect backend and database implementations.

Methodology

To bridge the gap between real-world web development and current agentic approaches, the authors introduce FullStack-Agent, a unified system comprising three tightly coupled components:

1. FullStack-Dev: A Multi-Agent Development Framework

FullStack-Dev orchestrates the full-stack generation workflow through a multi-agent system inspired by real-world development processes:

• Planning Agent: Acts as a lead architect, generating high-level frontend and backend plans in JSON format. It defines page layouts, components, data flows, entities, and API endpoints with granular type definitions to ensure smooth data flow.
• Coding Agents: Two specialized agents (Frontend and Backend) execute the plans sequentially. They operate in sandbox environments and utilize a suite of tools including file reading/writing, shell execution, and code searching.
• Specialized Debugging Tools:
  • Frontend Debugging Tool: Automatically starts the website service and runs a GUI-agent debugging process. Unlike previous methods that rely on blind interaction, this tool monitors terminal and browser console outputs to dynamically generate test cases based on the current development state, efficiently localizing errors.
  • Backend Debugging Tool: Functions similarly to API debugging tools (e.g., Postman), automatically starting services, sending requests, and returning console outputs. This significantly reduces the number of iterations required for the Backend Coding Agent.

2. FullStack-Learn: Iterative Self-Improvement via Back-Translation

Recognizing that the backbone LLM's coding ability is crucial, FullStack-Learn employs a data-scaling and self-improvement pipeline:

• Repository Back-Translation: High-quality real-world repositories crawled from GitHub are transformed into agent trajectories. An Information Gathering Agent summarizes the repository structure and functionality. A Trajectory Back-Translation Agent then transcribes the existing repository into an empty template, guided by high-level plans. A rule-based program cleans these trajectories by removing traces of the original repository, resulting in synthetic development trajectories.
• Repository Augmentation: To scale data further, the system generates synthetic repositories by applying augmentation plans (simplification, extension, or application transition) to existing repositories. An Augmentation Implementing Agent executes these plans and verifies the results.
• Iterative Training: The system uses an iterative process where an initial model generates data, which is then used to train an intermediate model. This intermediate model scales the data via augmentation and back-translation, creating a larger dataset for final training, thereby enhancing the backbone LLM's full-stack capabilities without relying on stronger external models.

3. FullStack-Bench: Comprehensive Evaluation Benchmark

FullStack-Bench addresses the limitations of frontend-only evaluation by systematically testing three dimensions:

• Frontend Test: A GUI-agent judge evaluates UI interactions. Crucially, it appends a database interaction validation step, extracting logs to verify that frontend actions correspond to actual database operations.
• Backend Test: An agent gathers API endpoint information and validates functionality by sending requests to the backend, checking responses against expected outcomes.
• Database Test: The agent inspects database snapshots (schema and data rows) to verify if the required data structures and content exist.

Key Contributions

1. FullStack-Dev: A multi-agent framework with highly effective coding and debugging tools that significantly outperforms previous state-of-the-art methods in full-stack generation.
2. FullStack-Learn: An iterative self-improvement method that substantially enhances the backbone LLM's full-stack development abilities through repository augmentation and back-translation, enabling models to learn from real-world codebases.
3. FullStack-Bench: A novel benchmark that comprehensively evaluates frontend, backend, and database functionalities, moving beyond superficial UI checks to verify true full-stack implementation.

Experimental Results

Extensive experiments were conducted using Qwen3-Coder models as the backbone LLMs on the FullStack-Bench:

• FullStack-Dev Performance: Using Qwen3-Coder-480B-A35B-Instruct, FullStack-Dev achieved accuracies of 64.7% (Frontend), 77.8% (Backend), and 77.9% (Database). This represents an improvement of 8.7%, 38.2%, and 15.9% respectively over the previous state-of-the-art method (WebGen-Agent). Notably, FullStack-Dev achieved higher backend accuracy than frontend accuracy, indicating the generation of functional backends rather than mock data.
• FullStack-Learn Performance: Training Qwen3-Coder-30B-A3B-Instruct with the FullStack-Learn pipeline improved its accuracy by 9.7% (Frontend), 9.5% (Backend), and 2.8% (Database) across two rounds of self-improvement. This demonstrates that the method can effectively scale the capabilities of smaller models without relying on stronger teacher models.
• Ablation Studies: Removing the multi-agent mechanism or the specialized debugging tools resulted in significant performance degradation, confirming their critical roles. Similarly, training on trajectories generated via Repository Back-Translation yielded significantly better results than training on directly generated trajectories, validating the efficacy of using real-world repositories.

Significance and Claims

The paper claims that FullStack-Agent represents a significant step forward in agentic coding by addressing the specific complexities of full-stack development. The authors emphasize that their approach:

• Closes the Gap: It bridges the divide between real-world web development requirements and current agentic capabilities.
• Ensures Verifiability: By integrating database log verification into the evaluation pipeline, it prevents the "fake interactivity" common in current systems.
• Scalability: The FullStack-Learn method demonstrates that high-quality full-stack coding skills can be distilled into models through self-improvement on real-world data, offering a path to enhance models without requiring massive, expensive external datasets or stronger teacher models.

The authors conclude that their system comprehensively addresses the challenges of production-level full-stack development, offering a robust framework for generating verifiable, functional, and scalable web applications.