SCIENTIST - Research Agent

The SCIENTIST agent conducts research on the selected competition. It analyzes leaderboards, searches for relevant papers and notebooks, and summarizes strategies.

Role

• Analyze leaderboard score distribution
• Surface relevant notebooks and research
• Summarize dataset characteristics
• Recommend modeling approaches

Tools

Tool	Purpose
analyze_leaderboard	Summarize leaderboard stats
get_kaggle_notebooks	Find top notebooks for the competition
analyze_data_characteristics	Inspect dataset structure
compute_baseline_estimate	Estimate a baseline score
kaggle_* toolset	Kaggle API helper tools
web_search (builtin)	Retrieve papers and discussions
memory (builtin)	Shared notes (Anthropic only)

Basic Usage

from agent_k.agents.scientist import ScientistDeps, scientist_agent

competition = await kaggle_adapter.get_competition("titanic")

deps = ScientistDeps(
    http_client=http,
    platform_adapter=kaggle_adapter,
    competition=competition,
)

run_result = await scientist_agent.run(
    "Research the provided competition and summarize approaches",
    deps=deps,
)

output = run_result.output
print(output.recommended_approaches)

Dependencies

from dataclasses import dataclass, field
from typing import Any
import httpx

@dataclass
class ScientistDeps:
    """Dependencies for the SCIENTIST agent."""

    http_client: httpx.AsyncClient
    platform_adapter: PlatformAdapter
    competition: Competition
    leaderboard: list[LeaderboardEntry] = field(default_factory=list)
    research_cache: dict[str, Any] = field(default_factory=dict)

Output Model

class ResearchFinding(BaseModel):
    """Individual research finding."""

    category: str
    title: str
    summary: str
    relevance_score: float
    sources: list[str]

class LeaderboardAnalysis(BaseModel):
    """Leaderboard statistics summary."""

    top_score: float
    median_score: float
    score_distribution: str
    common_approaches: list[str]
    improvement_opportunities: list[str]

class ResearchReport(BaseModel):
    """Output from SCIENTIST research."""

    competition_id: str
    domain_findings: list[ResearchFinding]
    technique_findings: list[ResearchFinding]
    leaderboard_analysis: LeaderboardAnalysis | None
    recommended_approaches: list[str]
    estimated_baseline_score: float | None
    key_challenges: list[str]

Notes

• The SCIENTIST output is converted into a simplified ResearchFindings object by the mission graph.
• The memory tool is only available for Anthropic models.