Guide

Guides

Pick the task you are doing. Each guide is self-contained and links to the concept or reference it relies on.

Build a head

• Onboard from logs

  Turn an agent's logs into a head with tabagent onboard. Privacy and cost controls, LLM providers, stopping and resuming.

• Bring your own table

  The scikit-learn estimators on a decision table you already have. No LLM.

• Choose and evaluate a head

  Metrics per decision kind, task-grouped cross-validation, paired intervals, other classifiers, discovery inside each fold.

Check and debug

• Check and debug logs

  tabagent scan, the supported log formats, how files are read, and tabagent check-logs.

• When onboarding goes wrong

  The checks that stop a run early and what to change when discovery fails at a stage.

• Verify the generated code

  The four value checks, the golden tests, and how to review extract.py before you serve it.

Deploy

• Deploy and serve

  Saved heads, live prefixes, the HTTP service, the sandbox, redaction and pretrained heads on the Hub.

• Hot-swap into an agent

  ToolFilter in front of a tool-calling LLM, the CUGA helpers, and whether filtering the tool list pays.

• Production runbook

  Scale, the warm extractor pool, timeouts, metrics and alerts, drift, upgrading and the security model.

Reproduce

• Reproduce the paper

  Every public result, the command that recomputes it, and the paper's value next to the released code's.

• Public datasets

  Each dataset's pinned source, license and the paper result it backs.