Get started

Core concepts

Five ideas, each with a page of its own in the Concepts section.

A closed-set decision

An agent repeatedly makes decisions whose answer comes from a fixed set: shortlist tools or skills, pick an application or an operation, route to a model or a sub-agent, activate a skill or not. TabAgent replaces the LLM call that makes such a decision. It does not generate free text or plan. If a component writes free text, it is not a candidate; tabagent scan says so.

One row per (context, candidate)

TabAgent casts every closed-set decision as one pointwise problem: a row is a (decision context, candidate) pair, its columns are behavior-grounded features, and its label says whether the agent's successful run chose that candidate. A head scores every candidate of a decision in one pass.

Kind	Example	Decoding
ranking	tool, skill, API, GUI-element shortlisting	top-k by score
multilabel	application selection	score above a threshold, back off to the best one
single	operation type, task category	argmax class
policy	model routing, sub-agent routing, skill gating	argmax predicted success

Decision kinds has the metrics and the decision table.

Three phases

TabSchema

Discovers features. LLM agents analyze the successful trajectories, write and sandbox-test extraction code, and a judge ensemble keeps the features that matter.

TabSynth

Augments the rows. Deterministic, label-preserving paraphrases of the free-text state, where trajectories are scarce.

TabHead

Decides in milliseconds. A CatBoost classifier scores every candidate in one pass.

Each phase runs once per agent and catalog. At inference the head replaces the LLM call; the rest of the agent is unchanged. The three phases goes through each.

The feature card

The tabular layout TabAgent learns from, with one row per (context, candidate) pair and one column per feature. Features come in three families, as in the paper's Fig. 1: schema, state and dependency. Dependency features are fitted statistics (a candidate's prior, co-usage rates, embedding similarity, k-NN outcomes), fitted on the training rows only. A card is saved as YAML (card.yaml) next to the head. Per-candidate features covers features that relate the context to a candidate.

The confidence gate

A head's confidence says how sure it is. A decision below a threshold is deferred: the agent calls the original LLM instead. The confidence gate explains how the threshold is calibrated and what it costs.

Leakage safety

Every data-dependent stage is fitted inside the training portion of each cross-validation fold, so a reported number is an estimate for new tasks. Leakage safety says how, and where the one exception is.