Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces FullStack-Agent, a multi-agent full-stack web coding system (FullStack-Dev) with debugging tools, a repository back-translation and augmentation self-improvement method (FullStack-Learn), and a full-stack benchmark (FullStack-Bench) testing frontend, backend, and database.

Problem

Existing LLM code agents generate mostly frontend-only websites lacking real backend and database functionality, and existing benchmarks judge only GUI-level interactions, missing backend and database correctness.

Approach

Multi-agent planning/coding with frontend and backend debugging tools; training data is produced by back-translating crawled GitHub website repositories (and augmented repositories) into agent trajectories, and the model is iteratively self-improved; the benchmark adds backend API and database test cases judged by agents plus database-log validation.

Key results

• FullStack-Dev with Qwen3-Coder-480B reaches 64.7% FE, 77.8% BE, and 77.9% DB accuracy, beating WebGen-Agent by 8.7, 38.2, and 15.9 points.
• FullStack-Learn two rounds raise Qwen3-Coder-30B-A3B by 9.7 (FE), 9.5 (BE), and 2.8 (DB) points on FullStack-Bench.
• Human alignment of the automatic judge is 90.5% (frontend), 94.0% (backend), and 97.5% (database) on 200 sampled instances each.

Surprising findings

• Directly generated trajectories from user instructions did not notably improve the 30B model, while back-translated repository trajectories gave large gains.
• Baseline agents' backend accuracy is much lower than frontend accuracy, often relying on mock data.

Limitations

• Benchmark has only 101 user instructions and relies on LLM-annotated and human-refined test cases.
• Evaluation uses only two templates (Next.js and NestJS) in main experiments.
• Baselines were prompted to generate backend components, so comparisons depend on the prompt wrapping.

Experimental setup

Datasets: crawled GitHub website repositories WebGen-Bench user instructions WebGen-Instruct

Benchmarks: FullStack-Bench WebGen-Bench

Metrics: frontend accuracy frontend accuracy with valid database backend accuracy backend accuracy with valid database database accuracy appearance score

Baselines: WebGen-Agent TDDev OpenHands Bolt.diy Qwen-Code

Builds on: WebGen-Bench WebGen-Agent OpenHands Qwen3-Coder

Closest papers by meaning

• WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning (ICLR 2026 · similarity 0.89)
• WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch (NeurIPS 2025 · similarity 0.89)
• Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification (ICML 2026 · similarity 0.88)
• NL2Repo-Bench: Towards Long-Horizon Repository Generation Evaluation of Coding Agents (ICML 2026 · similarity 0.87)
• GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging (AAAI 2026 · similarity 0.87)
• EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · similarity 0.87)