Weco: The Code Optimization Agent

pip install weco

Weco systematically optimizes your code, guided directly by your evaluation metrics.

Example applications include:

• GPU Kernel Optimization: Reimplement PyTorch functions using CUDA or Triton, optimizing for latency, throughput, or memory_bandwidth.
• Model Development: Tune feature transformations, architectures or the whole training pipeline, optimizing for validation_accuracy, AUC, or Sharpe Ratio.
• Prompt Engineering: Refine prompts for LLMs (e.g., for math problems), optimizing for win_rate, relevance, or format_adherence

Overview

The weco CLI leverages a tree search approach guided by LLMs to iteratively explore and refine your code. It automatically applies changes, runs your evaluation script, parses the results, and proposes further improvements based on the specified goal.

Install the Package

pip install weco

Getting Started

Quickstart with an example project

Configure optimization parameters yourself - If you need precise control over the optimization parameters, you can use the direct weco run command:

Example: Optimizing Simple PyTorch Operations

git clone https://github.com/WecoAI/weco-cli.git
cd weco-cli/examples/hello-world/
pip install -r requirements.txt

# Run Weco with configuration
weco run --source module.py \
     --eval-command "python evaluate.py --path module.py" \
     --metric speedup \
     --goal maximize \
     --steps 10 \
     --additional-instructions "Fuse operations in the forward method while ensuring the max float deviation remains small. Maintain the same format of the code."

Note: If you have an NVIDIA GPU, change the device in the --eval-command to cuda. If you are running this on Apple Silicon, set it to mps.

For more advanced examples, including Triton, CUDA kernel optimization, ML model optimization, and prompt engineering for math problems, please see the README.md files within the corresponding subdirectories under the examples/ folder.

Note: When recommend removing any backticks from your code if any are present. We currently don't support backticks but will support this in the future.

Arguments for weco run

Required:

Optional:

Command Reference

Basic Usage Patterns

Setup Commands (Experimental)

The setup command installs Weco skills for AI coding assistants:

weco setup claude-code  # For Claude Code
weco setup cursor       # For Cursor

• Claude Code: Downloads the Weco skill to ~/.claude/skills/weco/ and updates ~/.claude/CLAUDE.md
• Cursor: Downloads the Weco skill to ~/.cursor/skills/weco/ and creates ~/.cursor/rules/weco.mdc

Model Selection

You can specify which LLM model to use with the -M or --model flag:

weco run --model gpt-5 --source optimize.py [other options...]

Available models (31 total):

OpenAI Models:

• GPT-5 Series: gpt-5.2, gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini, gpt-5-codex, gpt-5-pro, gpt-5, gpt-5-mini, gpt-5-nano
• O-Series Reasoning: o3-pro, o3, o3-mini, o4-mini, o1-pro, o1, codex-mini-latest
• GPT-4 Series: gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini

Anthropic Claude (via Vertex AI):

• claude-opus-4-5, claude-opus-4-1, claude-opus-4, claude-sonnet-4-5, claude-sonnet-4, claude-haiku-4-5

Google Gemini:

• gemini-3-pro-preview, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite

All models are available through Weco. If no model is specified, Weco automatically selects the best model for your optimization task.

Resuming Interrupted Runs

If your optimization run is interrupted (network issues, restart, etc.), resume from the most recent node:

# Resume an interrupted run
weco resume 0002e071-1b67-411f-a514-36947f0c4b31

Arguments for weco resume:

Notes:

• Works only for interrupted runs (status: error, terminated, etc.).
• You’ll be prompted to confirm that your evaluation environment (source file + evaluation command) hasn’t changed.
• The source file is restored to the most recent solution before continuing.
• All progress and metrics from the original run are preserved.
• Log directory, save-logs behavior, and evaluation timeout are reused from the original run.

Performance & Expectations

Weco, powered by the AIDE algorithm, optimizes code iteratively based on your evaluation results. Achieving significant improvements, especially on complex research-level tasks, often requires substantial exploration time.

The following plot from the independent Research Engineering Benchmark (RE-Bench) report shows the performance of AIDE (the algorithm behind Weco) on challenging ML research engineering tasks over different time budgets.

As shown, AIDE demonstrates strong performance gains over time, surpassing lower human expert percentiles within hours and continuing to improve. This highlights the potential of evaluation-driven optimization but also indicates that reaching high levels of performance comparable to human experts on difficult benchmarks can take considerable time (tens of hours in this specific benchmark, corresponding to many --steps in the Weco CLI). Factor this into your planning when setting the number of --steps for your optimization runs.

Saving Execution Logs

When using the --save-logs flag, Weco saves the execution output from each optimization step to help with debugging and analysis. The logs are organized as follows:

.runs/
└── <source-file-name>/
    └── <run-uuid>/
        ├── exec_output.jsonl      # Index file with metadata for each step
        ├── outputs/
        │   ├── step_0.out.txt      # Raw output from initial evaluation
        │   ├── step_1.out.txt      # Raw output from step 1
        │   ├── step_2.out.txt      # Raw output from step 2
        │   └── ...
        ├── step_0.py               # Code snapshot from initial evaluation
        ├── step_1.py               # Code snapshot from step 1
        ├── step_2.py               # Code snapshot from step 2
        └── ...

Each run is organized under the source file name (e.g., spaceship-titanic for spaceship-titanic.py) and a unique UUID. The outputs/ directory and exec_output.jsonl file are only created when the --save-logs flag is used.

The exec_output.jsonl file contains one JSON object per line with:

• step: The optimization step number
• timestamp: When the execution occurred
• output_file: Relative path to the full output file
• output_length: Total length of the output

This is particularly useful for:

• Debugging why certain optimizations fail
• Analyzing patterns in evaluation results
• Keeping records of long-running optimization sessions
• Troubleshooting evaluation script issues

Important Note on Evaluation

The command specified by --eval-command is crucial. It's responsible for executing the potentially modified code from --source and assessing its performance. This command MUST print the metric you specified with --metric along with its numerical value to the terminal (standard output or standard error). Weco reads this output to understand how well each code version performs and guide the optimization process.

For example, if you set --metric speedup, your evaluation script (eval.py in the examples) should output a line like:

speedup: 1.5

or

Final speedup value = 1.5

Weco will parse this output to extract the numerical value (1.5 in this case) associated with the metric name ('speedup').

Note on Output Truncation: When evaluation output exceeds 51,000 characters, Weco truncates it to show the first 25,000 and last 25,000 characters. For best results, ensure your evaluation script prints the metric value near the end of its output.

Supported Models

A list of models we support can be found in our documentation here.

Contributing

We welcome contributions! Please see contributing.md for detailed guidelines on how to contribute to this project.