OPAL

A person can use OPAL to experiment with agent-adjusted optimizers during PyTorch training runs.

Availability

Install with `pip install opal-optimizer`; source and guide are on GitHub

Pricing model

Agent usage is billed to the Claude subscription or API key; `max_cost_usd` can cap spending

Deployment

Runs inside the training process on CPU, one GPU, or multiple GPUs of one node (DDP)

Privacy

The agent can see the goal, parameter names and shapes, losses, analysis reports, and read-only training script and model code; `share_code=False` keeps code out

Licence

Not established

• OPAL lets a coding agent rewrite optimizer and analysis programs during a training run. ^[1]

  rewrite two small programs: the optimizer program, which turns gradients into updates at every step, and an analysis

• The package is available on PyPI as `opal-optimizer`. ^[1]

  The package is called `opal-optimizer` on PyPI, because `opal` there is an unrelated web framework.

• OPAL supports PyTorch training on CPU, one GPU, or several GPUs of one node (DDP). ^[1]

  - **A PyTorch training script** that uses `torch.optim` optimizers, on CPU, one GPU or several GPUs of one node (DDP).

• Agent usage is billed to the Claude subscription or API key, and there is no cost cap by default. ^[1]
• The agent can receive read-only copies of the training script and model code; `share_code=False` keeps them out. ^[1]

  analysis reports, and read-only copies of your training script and model code (`share_code=False` keeps them out).

• A speech model adds multilingual transcription and translation Daily brief 2026-10-26