FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation

• Natural-language coding agents often produce frontend-only sites that simulate interactivity without backend processing or storage, while production-grade full-stack development requires code navigation, long-term reasoning, and evaluation beyond UI checks.Source: §1 · p. 1
• FullStack-Dev assigns architecture and planning to a lead agent, delegates frontend and backend work to coding agents, and equips them with editing, command-execution, and debugging tools.Its debugging tools are designed to locate and correct subtle errors.Source: §1 · p. 2
• FullStack-Learn augments and back-translates GitHub website repositories into supervised fine-tuning trajectories, combining global planning and information gathering with local code implementation.The trajectories aim to teach agentic coding abilities and framework expertise from real-world codebases.Source: §1 · p. 2
• FullStack-Bench tests frontend, backend, and database functionality with agent judges and evaluates database logs to check whether website actions involve adequate database interactions.Source: §1 · p. 2
• FullStack-Dev with Qwen3-Coder-480B-A35B-Instruct achieves 64.7% frontend, 77.8% backend, and 77.9% database accuracy, beating the previous state of the art by 8.7%, 38.2%, and 15.9%.Source: §1 · p. 2
• FullStack-Learn improves Qwen3-Coder-30B-Instruct accuracy by 9.7% on frontend, 9.5% on backend, and 2.8% on database tests.Source: §1 · p. 2

2.1. FullStack-Dev

• FullStack-Dev plans frontend and backend work, builds the backend before the frontend, and uses sandboxed agents and debugging tools to control development.Its GUI debugging tool monitors browser and terminal errors and generates tests dynamically; its backend tool runs API requests and returns responses and console output.Source: §2.1 · pp. 2–3

2.2. FullStack-Learn

• Repository Back-Translation turns existing repositories into development-like agent trajectories, using repository summaries and plans to guide reconstruction in an empty framework template.A quality score filters low-quality repositories, and a rule-based program removes traces of the original repository while preserving trajectory coherence.Source: §2.2 · p. 4
• Repository Augmentation creates five times as many repositories by implementing planned changes, then debugs and verifies augmented samples before retaining them.The plans include simplification, extension, and application transitions; the resulting repositories expand the trajectories available for training.Source: §2.2 · p. 4
• Iterative self-improvement trains an intermediate model on back-translated real repositories, then trains the final model on those trajectories plus data from augmented repositories.The initial dataset is small because of the original model’s limited full-stack coding ability; the intermediate model generates trajectories from augmented repositories for the expanded dataset.Source: §2.2 · pp. 4–5

2.3. FullStack-Bench

• FullStack-Bench evaluates websites from natural-language instructions with 647 frontend, 604 backend, and 389 database test cases across 101 instructions.The frontend cases come from WebGen-Bench; backend and database cases were constructed with LLM annotation and human refinement. Frontend judgments require database-interaction checks, while backend API requests and database snapshots assess their respective test cases.Source: §2.3 · p. 5

3.1. FullStack-Dev Results

• 64.7%, 77.8%, and 77.9%: FullStack-Dev with Qwen3-Coder-480B-A35B-Instruct beats WebGen-Agent by 8.7%, 38.2%, and 15.9% in frontend, backend, and database accuracy.With Qwen3-Coder-30B-A3B-Instruct, FullStack-Dev also achieves the highest accuracies: 37.2% frontend, 38.7% backend, and 50.9% database. It also achieves the highest appearance score.Source: §3.1 · p. 6, Table 2 · p. 6
• FullStack-Dev’s backend accuracy exceeds its frontend accuracy, while most baseline methods show the reverse, indicating that its generated sites generally have functional backends.The authors attribute many baselines’ lower backend accuracy to their reliance on mock data despite explicit requests to generate backend components.Source: §3.1 · p. 6

3.2. FullStack-Learn Results

• 46.9%, 48.2%, and 53.7% frontend, backend, and database accuracy: two-round FullStack-Learn improves over the original 30B model by 9.7%, 9.5%, and 2.8%.The process is self-improving without a stronger model: round one uses 2K trajectories, and round two adds 8K generated from augmented repositories.Source: §3.2 · pp. 6–7, Table 3 · p. 7

3.3. Ablation Studies

• Removing the multi-agent mechanism reduces accuracy across all metrics; removing either debugging tool hurts its corresponding frontend or backend accuracy more, and removing both degrades both.The reported effects align with the tools’ respective roles in frontend and backend debugging.Source: §3.3 · p. 7
• Repository Back-Translation trajectories significantly improve accuracies and appearance, whereas directly generated trajectories from WebGen-Instruct fail to notably improve overall performance.The comparison uses 2K trajectories for each approach; the authors suggest the gap stems from learning from high-quality real-world repositories.Source: §3.3 · p. 7
• Human alignment accuracy exceeds 90% for each of FullStack-Bench’s frontend, backend, and database tests, supporting the reliability of its testing pipeline.Four student volunteers checked 200 randomly sampled instances from each test area, counting a sample as correct only when its evaluation trajectory and database logs supported the result.Source: §3.3 · p. 7

Website Development Agents and Pipelines

• Many website agents generate frontend-only or simple sites and lack dynamic code navigation; FullStack-Dev supports dynamic tests and error localization.WebGen-Agent and TDDev also use GUI feedback, but their test cases are predefined and their agents interact blindly (Lu et al., 2025a; Wan et al., 2025).Source: §4 · p. 7

Website Development Benchmarks

• FullStack-Bench tests frontend, backend, and database implementations, addressing benchmarks that rely on fixed pipelines or GUI-agent results alone.Some benchmarks evaluate simple HTML generation from design images, while WebGen-Bench evaluates multi-file agentic coding but judges websites solely through GUI interactions (Si et al., 2025; Yun et al., 2024; Guo et al., 2025; Xiao et al., 2025b;a; Sun et al., 2025; Zhang et al., 2025a; Xu et al., 2025; Lu et al., 2025b).Source: §4 · p. 8

Training Methods to Improve Software Development

• FullStack-Learn back-translates augmented repositories into self-improvement trajectories without relying on stronger models.This differs from screenshot–HTML SFT and GitHub-issue patch training, which target other tasks (Gui et al., 2025; Laurenc¸on et al., 2024; Wei et al., 2025; Ma et al., 2025; Zhang et al., 2025b; Yang et al., 2025; Pan et al., 2025). WebGen-Bench uses larger LLMs for synthetic trajectories, while WebGen-Agent uses GRPO with visual feedback but does not leverage existing codebases (Lu et al., 2025a).Source: §4 · p. 8

5. Conclusion

• FullStack-Dev with Qwen3-Coder-480B-A35B-Instruct achieves 64.7% frontend, 77.8% backend, and 77.9% database accuracy, exceeding the previous state-of-the-art method by 8.7%, 38.2%, and 15.9%, respectively.These results are on FullStack-Bench.Source: §5 · p. 8
• Training Qwen3-Coder-30B-Instruct with FullStack-Learn improves accuracy by 9.7% on frontend, 9.5% on backend, and 2.8% on database test cases.The improvements are reported for the three FullStack-Bench test sets, respectively.Source: §5 · p. 8
• FullStack-Agent unifies FullStack-Dev's multi-agent development framework and coding and debugging tools, FullStack-Learn's repository augmentation and back-translation, and FullStack-Bench's evaluation of frontend, backend, and database functionality.Source: §5 · p. 8

Impact Statement

• The work may reduce demand for human programmers, while unclear responsibility for errors or vulnerabilities in agent-generated websites could create new challenges.The authors highlight these consequences to emphasize the broader societal impact of their work and the importance of responsible deployment.Source: Impact Statement · p. 9

A. Tool Details

• Code-reading and navigation tools include reading one or many files, listing directories, globbing for files, and searching file contents.Source: A. Tool Details · p. 11
• Code-editing tools let the agent write files and replace strings, while a shell-command tool executes commands.Source: A. Tool Details · p. 11
• Debugging tools test backend APIs and the frontend.Source: A. Tool Details · p. 11

B. Increasing the Number of Templates

• The experiments use Next.js2 and NestJS3 as frontend and backend templates to stabilize ablation comparisons and reduce data-generation and training costs.The authors state that this choice does not affect the conclusions drawn from the experimental results.Source: B. Increasing the Number of Templates · pp. 11–12
• Adding Django and Vue.js slightly increased frontend, backend, and database test accuracies, supporting FullStack-Dev’s ability to generalize across website-development templates.The authors suggest that more template choices may help the agent find suitable, easier-to-work-with options.Source: Table 8 · p. 12, B. Increasing the Number of Templates · p. 12
• FullStack-Dev can choose frontend and backend templates based on the user instruction and template descriptions.The expanded set adds Django4 and Vue.js5 to the original Next.js2 and NestJS3 templates.Source: B. Increasing the Number of Templates · p. 12, Table 9 · p. 12

C. Back-Translation Trajectory Transforming Process

• The resulting trajectories follow the high-level plans and use professional, direct, concise agent messages, aligning them with normal coding trajectories.The message tone is requested by the system prompt.Source: C. Back-Translation Trajectory… · p. 12
• The process rewrites paths to an adjusted project, canonicalizes staged prompts from the instruction, relevant plans, and scaffold, and removes old-repository references and dependent tool calls.It replaces staged implementation or validation messages with canonical coding-agent prompts and removes associated tool responses when calls depend on paths outside the new trajectory.Source: C. Back-Translation Trajectory… · p. 13
• After resetting the project, the process replays state-mutating actions, recomputes outputs for inspection calls, and replaces the original tool responses with corrected outputs.The replay follows chronological order in a tool runtime bound to the adjusted project path.Source: C. Back-Translation Trajectory… · p. 13

D. Data Filtering Details

• Frontend appearance and functionality scores range from 1 to 5, while backend scores are 1 for nonempty HTTP 200 responses, 0 for empty HTTP 200 responses, and -1 otherwise.The frontend scores are extracted from tool-call summaries; backend scores are derived from response codes and data.Source: D. Data Filtering Details · p. 13
• Aggregated scores use γ = 0.9, giving earlier tool-call scores lower weight; positive signals are defined by thresholds of 3 for frontend scores and 0 for backend scores.The threshold separates positive from negative signals for each score type.Source: D. Data Filtering Details · pp. 13–14
• A trajectory is kept only when the aggregated appearance, frontend functionality, and backend functionality scores are all above zero.Source: D. Data Filtering Details · p. 14

E. Baseline Implementation Details

• Website-development and general coding agents served as baselines, with prompts explicitly requesting backend and database components when appropriate.The agents typically generated frontend-only pages by default.Source: E. Baseline Implementation Details · p. 14

1. WebGen-Agent: https://github.com/mnluzimu/WebGen-Agent

• WebGen-Agent (Lu et al., 2025a) was included among the website-development agent baselines.Source: E. Baseline Implementation Details · p. 14

5. Qwen-Code: https://github.com/QwenLM/qwen-code

• Qwen-Code (Qwen, 2025b) was included among the general coding agent baselines.Source: E. Baseline Implementation Details · p. 14

E. Baseline Implementation Details

• The baseline agents’ open-source implementations were used to generate websites for FullStack-Bench.Source: E. Baseline Implementation Details · p. 14

5. Qwen-Code: https://github.com/QwenLM/qwen-code

• Baseline implementation code will also be released to support reproducibility.Source: §5 · p. 14

F. Error Analysis

• Frontend errors were led by “Functionality Not Implemented” (29.7%) and “Unresponsive Components” (23.7%); “Website Start Failed” accounted for 14.3%, while data-fetching and form-submission errors each accounted for 9.7%.The remaining small-number categories included Incorrect Color Theme, Referring to Missing Modules, and Syntax Error.Source: F. Error Analysis · pp. 15–16
• Backend errors most often involved No Database Interaction (34.3%) or API Not Implemented (33.3%); database setup and connection failures accounted for 19.7% and 12.7%.No Database Interaction includes fake responses without database reads or writes.Source: F. Error Analysis · p. 16
• Database errors were led by Database Empty (46.7%), followed by Data Fields Missing (26.0%), Tables Missing (19.7%), and Data Structure Not Clearly Provided (7.7%).The analysis attributes many database errors to failure to initialize the database.Source: F. Error Analysis · p. 16

G. Human Annotation Details

• Four student volunteers manually checked 200 frontend, 200 backend, and 200 database instances; correctness required trajectories and database logs to support the final result.Human alignment accuracy is the percentage of correct samples.Source: G. Human Annotation Details · p. 16, Figure 4 · p. 17, Figure 5 · p. 17

H. FullStack-Dev Prompts

• FullStack-Dev is template-agnostic, but experiments use Next.js and NestJS; a planning agent produces frontend and backend plans that seed the coding agents.Template selection requires a name, description, and guideline.Source: H. FullStack-Dev Prompts · p. 16, Figure 8 · p. 20, Figure 9 · p. 21
• FullStack-Dev prompts define agent mandates and workflows, with starting, validation, and summary prompts for backend coding and starting and validation prompts for frontend coding.The system prompt specifies operational constraints.Source: H. FullStack-Dev Prompts · p. 17, Figure 10 · p. 22, Figure 11 · p. 23
• The Frontend Debugging Tool’s GUI-agent summary records the test trajectory, the action triggering an error or misbehavior, website appearance, and functionality and appearance scores.It handles natural completion and premature termination due to runtime errors differently.Source: H. FullStack-Dev Prompts · p. 17, J. FullStack-Bench Prompts · p. 25, Figure 13 · p. 25

I. FullStack-Learn Prompts

• FullStack-Learn uses prompts for repository information gathering, trajectory back-translation, and augmentation planning and implementation, retaining only augmentations that pass verification.Verification checks whether the augmentation plan was fully implemented using concrete code evidence.Source: §I · p. 18, Figure 14 · p. 26, Figure 15 · p. 27

J. FullStack-Bench Prompts

• FullStack-Bench prompts test frontend behavior with database-interaction validation, gather backend API information, test backend APIs with database validation, check database contents, and grade appearance from screenshots on a 1–5 scale.The appearance-grading prompt is adapted from WebGen-Bench.Source: J. FullStack-Bench Prompts · p. 18, Figure 18 · p. 30, Figure 19 · p. 31