DeepBuilder

Introduction

DeepBuilder is an AI agent that automatically builds AI models. Given a task, it runs an agent loop that analyzes the problem, designs models, writes code to implement them, trains them, tunes hyperparameters, evaluates model performance, and iteratively improves the models. By automating the model development workflow, DeepBuilder reduces much of the manual effort required to build AI models.

Key Technical Contributions

Multi-Mode Planning & Code Generation — Supports base (single-shot) and memory-enhanced (two-stage retrieval-augmented) planning, paired with three code generation strategies: single-pass, stepwise multi-agent pipeline, and incremental search/replace diff patching. Different modes are dispatched adaptively based on search state.

Experience-Driven Memory — A global memory layer records plan, code, metrics, and success/failure labels for every node. Retrieval combines BM25 + FAISS allowing the planner to reinforce proven strategies and avoid known pitfalls from its own search history. Different agents query memory in different ways to encourage novel approaches.

Progressive MCGS (Monte Carlo Graph Search) with Cross-Branch Fusion — The search graph extends vanilla UCT with piecewise exploration decay, time-aware explore-exploit switching, and automatic stagnation detection. Multiple solution branches evolve in parallel; when progress stalls, the system performs cross-branch fusion — merging insights from top-performing nodes across different branches into new solution candidates — and trajectory-aware evolution that leverages each branch’s full improvement history to propose informed next steps.

Current Results

On OpenAI MLE-bench, DeepBuilder ranked #1, demonstrating strong performance on real-world AI model building tasks.