research

agentic_ml.research

The research orchestrator and agent glue (requires the research extra).

Researcher

Researcher(task: Task, model: object | None = None)

Run agent-driven research for a given task and export the best model.

Attributes:

Name	Type	Description
task		The :class:~agentic_ml.core.task.Task defining the problem being solved (prompt sections, metrics, partitioning and evaluation scheme).
model		The Strands model instance driving the agent, or None if this researcher was built without one — in that case :meth:research raises a :class:ValueError as soon as it is called.
agent	object | None	The Strands agent built from task/model, created lazily the first time :meth:research runs; None before that.
context	ResearchContext	The :class:~agentic_ml.core.engine.ResearchContext from the last :meth:research call; raises :class:RuntimeError if accessed before research has run.

Source code in src/agentic_ml/research/researcher.py

def __init__(self, task: Task, model: object | None = None) -> None:
    self.task = task
    self.model = model
    self._ctx: ResearchContext | None = None
    self.agent: object | None = None

context property

context: ResearchContext

The :class:~agentic_ml.core.engine.ResearchContext from the last research run.

research

research(
    data: DataFrame,
    schema: DatasetSchema,
    *,
    metrics: list[str] | None = None,
    partitions: list[Partition] | None = None,
    n_splits: int = 5,
    seed: int = 0,
    tolerance: float = 1e-06,
    timeout: float = 120.0,
    iterations: int = 10,
    ideas: str | None = None,
    research_dir: str | None = None,
    sleep: float = 0.0,
    new_session: bool = True,
    agent_config: AgentConfig | None = None,
) -> ResearchContext

Drive the research loop and return the resulting context.

Parameters:

Name	Type	Description	Default
data	DataFrame	The tabular dataset, including the target column.	required
schema	DatasetSchema	Variable descriptions and the target column name.	required
metrics	list[str] | None	Metric names to track; defaults to the task's metrics.	None
partitions	list[Partition] | None	Optional explicit (train_idx, val_idx) splits; otherwise built by the task (K-Fold by default).	None
iterations	int	Hard cap on the number of trials create_trial will actually create, enforced by create_trial_action (not just suggested to the agent via the kickoff prompt, which also mentions this number so the agent can plan ahead).	10
ideas	str | None	Natural-language ideas the user wants the agent to try.	None
research_dir	str | None	Where to store trials and the leaderboard. Defaults to a timestamped directory under ./agentic_ml_runs.	None
sleep	float	How long to wait between iterations, in seconds. This is useful to avoid RateLimit errors with the LLM provider.	0.0
new_session	bool	If True, clears any previous session cache under research_dir and starts a brand-new agent. If False, rebuilds the agent from scratch against this call's fresh context but restores its conversation from the on-disk session cache (whether written by an agent still alive on this instance or by an earlier process that has since restarted). Only meaningful when calling :meth:research again with the same research_dir as a previous call.	True
agent_config	AgentConfig | None	Tunes the Strands agent's conversation manager and context offloader (see :class:~agentic_ml.research.agent.AgentConfig); defaults to AgentConfig() (all field defaults) when not passed.	None

Source code in src/agentic_ml/research/researcher.py

def research(
    self,
    data: pd.DataFrame,
    schema: DatasetSchema,
    *,
    metrics: list[str] | None = None,
    partitions: list[Partition] | None = None,
    n_splits: int = 5,
    seed: int = 0,
    tolerance: float = 1e-6,
    timeout: float = 120.0,
    iterations: int = 10,
    ideas: str | None = None,
    research_dir: str | None = None,
    sleep: float = 0.0,
    new_session: bool = True,
    agent_config: AgentConfig | None = None,
) -> ResearchContext:
    """Drive the research loop and return the resulting context.

    Args:
        data: The tabular dataset, including the target column.
        schema: Variable descriptions and the target column name.
        metrics: Metric names to track; defaults to the task's metrics.
        partitions: Optional explicit (train_idx, val_idx) splits; otherwise built by the
            task (K-Fold by default).
        iterations: Hard cap on the number of trials ``create_trial`` will actually create,
            enforced by ``create_trial_action`` (not just suggested to the agent via the
            kickoff prompt, which also mentions this number so the agent can plan ahead).
        ideas: Natural-language ideas the user wants the agent to try.
        research_dir: Where to store trials and the leaderboard. Defaults to a
            timestamped directory under ``./agentic_ml_runs``.
        sleep: How long to wait between iterations, in seconds. This is useful
            to avoid RateLimit errors with the LLM provider.
        new_session: If ``True``, clears any previous session cache under ``research_dir``
            and starts a brand-new agent. If ``False``, rebuilds the agent from scratch
            against this call's fresh context but restores its conversation from the
            on-disk session cache (whether written by an agent still alive on this instance
            or by an earlier process that has since restarted). Only meaningful when
            calling :meth:`research` again with the *same* ``research_dir`` as a previous
            call.
        agent_config: Tunes the Strands agent's conversation manager and context offloader
            (see :class:`~agentic_ml.research.agent.AgentConfig`); defaults to
            ``AgentConfig()`` (all field defaults) when not passed.
    """
    if self.model is None:
        raise ValueError(
            "a Strands model instance is required to run research; pass one via "
            "Researcher(model=...) (install the 'research' extra)"
        )

    from agentic_ml.research.agent import AgentConfig as _AgentConfig

    agent_config = agent_config or _AgentConfig()

    chosen_metrics = metrics or self.task.default_metrics()
    primary = self.task.primary_metric()
    if primary not in chosen_metrics:
        chosen_metrics = [primary, *chosen_metrics]

    config = ResearchConfig(
        metrics=chosen_metrics,
        primary_metric=primary,
        maximize=self.task.maximize(),
        n_splits=n_splits,
        seed=seed,
        tolerance=tolerance,
        timeout=timeout,
        iterations=iterations,
        ideas=ideas,
        partitions=partitions,
        sleep=sleep,
    )

    if research_dir is None:
        stamp = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
        research_dir = f"./agentic_ml_runs/{self.task.name}_{stamp}"

    ctx = setup_research(self.task, data, schema, config, research_dir, agent_config)
    self._ctx = ctx

    self._load_session(new_session)
    error = self._kickoff_research_agent(config)
    if error is not None:
        raise error

    return ctx

export_model

export_model(trial: str | None = None) -> AgenticModel

Export a trial (the best by default) as a light production model.

Source code in src/agentic_ml/research/researcher.py

def export_model(self, trial: str | None = None) -> AgenticModel:
    """Export a trial (the best by default) as a light production model."""
    ctx = self.context
    trial_id = trial or select_best(ctx)
    if trial_id is None:
        raise RuntimeError("no successful trial to export")
    return AgenticModel.from_trial(ctx.workspace.trial_dir(trial_id))