Data and research agents

Why it pays

• A full analyst's workstation Python, notebooks, a local Postgres in Docker and a browser, all on one machine.
• Reproducible Save the prepared environment as a template, so every run starts the same way.
• Hands-off Schedules start the work, webhooks deliver results, and budgets cap the spend.

Step by step

1. Prepare the workstation

   Start a database and install your libraries, then save it.

   CLI

   harebox new --name research
   harebox exec current -- "docker run -d --restart always -p 5432:5432 -e POSTGRES_PASSWORD=local postgres:17 && pip install pandas duckdb jupyterlab"
   harebox template save current --name research-v1

2. Run the analysis as an agent turn

   The agent reads data, writes code, runs it and writes a report.

   CLI

   harebox prompt current "Load sales.csv, find the three biggest changes this week, write report.md with charts" --harness claude --budget 3

   Python

   box = hb.sandbox(sandbox_id)
   for ev in box.prompt("Load sales.csv, find the three biggest changes this week, write report.md with charts",
                        harness="claude", budget_usd=3):
       pass
   report = hb.read_file(box.id, query={"path": "/home/user/report.md"})

3. Schedule it weekly

   Resume the sandbox each Monday and send the prompt.

   CLI

   harebox schedule new --body '{"action": "prompt", "cron": "0 6 * * 1", "target": "SANDBOX_ID", "payload": {"prompt": "Refresh report.md"}}'

Tips

• Keep raw data in the sandbox and export only results, so large files never cross your laptop.
• Use a budget on the research label, so a weekly job has a monthly ceiling.
• Open JupyterLab through a private endpoint when you want to look yourself.