Skip to main content

Task Submission Using AI Agents

Transformer Lab ships an Agent Skill that lets Claude Code (or any coding agent that supports skills) drive the lab CLI on your behalf. Once installed, you can ask your agent in natural language to create tasks, queue jobs on compute providers, stream logs, and pull down artifacts — without memorizing CLI flags or writing task.yaml by hand.

Prerequisites​

  • Install the lab CLI with uv:

    uv tool install transformerlab-cli

    See CLI for more install options and full command reference.

  • Claude Code (or another agent harness that supports the skills format) is installed.

The agent will walk you through logging in and picking a current experiment the first time it needs them.

Install the skill​

From your project directory, install the skill with:

npx skills add transformerlab/transformerlab-app --skill transformerlab-cli

This teaches your agent how to use lab to check job status, stream logs, download artifacts, queue tasks, manage providers, and more.

Example: create and queue a fine-tune from a prompt​

With the skill installed, start a Claude Code session and send:

create a task that finetunes SmolLM2 on a fake dataset using the huggingface library.

Claude will:

  • Scaffold a task directory with a task.yaml plus a small training script that uses the HuggingFace transformers library against a synthetic dataset.
  • Register it in your current experiment by running lab task add ./<task-dir>.
  • Report back the new task ID.

Once the task exists, follow up with:

ok now queue this on our skypilot provider

Claude will call lab task queue <task-id>, pick the SkyPilot provider, fill in parameter defaults from task.yaml, and hand back the job ID. You can then ask it to tail logs, list artifacts, or download results — all of which map to existing lab job ... subcommands.

What the agent can do​

The skill exposes the full lab surface area to the agent, including:

  • lab task add / task list / task info / task delete
  • lab task queue (with provider selection and parameter prompting)
  • lab job list / job info / job artifacts / job download
  • lab job monitor (interactive TUI)
  • lab config and lab status for inspecting CLI state

Because the agent is just driving the CLI, anything you can do from the terminal is available — the skill primarily teaches the agent when and how to reach for each command.

Where to go next​