Get started

Get started

From pip install to a decision head, in four short pages.

TabAgent replaces an LLM agent's repeated closed-set decisions (which tools to shortlist, which skill or application to use, which model to route to) with a cheap classifier trained on the agent's own successful runs. These four pages take you from nothing installed to a head you can serve.

• Installation

  pip install tabagent, the extras (embed, tabschema, llm, serve, hub) and the LLM keys.

• Quickstart (5 minutes)

  Rank tools from a table on CPU with no LLM and no network, then reproduce a paper number.

• Onboard your agent

  Scan the logs, check them, estimate the cost, onboard a decision and serve it.

• Core concepts

  Closed-set decisions, the decision table, the three phases and the confidence gate in one page.

Pick your path

• I have my decisions as a table

  Use the scikit-learn estimators: Bring your own table.

• I have an agent's logs

  Onboard from the logs: Onboard from logs.

• I want to check the paper

  Reproduce the paper from public data, with no LLM call.