Skip to content

AGENT-K

AGENT-K

Multi-Agent Kaggle Competition Framework

AGENT-K is an autonomous multi-agent system for discovering, researching, prototyping, evolving, and submitting solutions to Kaggle competitions. Built with Pydantic-AI and Pydantic-Graph, it orchestrates specialized agents through a 5-phase mission lifecycle.

Key Features

  • Multi-Agent Orchestration — Pydantic-Graph state machine coordinates specialized agents
  • Evolutionary Code Search — Population-based solution optimization with OpenEvolve
  • Kaggle Integration — FunctionToolset-based platform operations
  • Real-Time Observability — Pydantic Logfire instrumentation for tracing and metrics
  • Interactive Dashboard — Next.js frontend with mission monitoring

The Agents

Agent Role Description
LYCURGUS Orchestrator Coordinates the multi-agent lifecycle using a state machine
LOBBYIST Discovery Finds competitions matching criteria via Kaggle API + web search
SCIENTIST Research Analyzes leaderboards, papers, and winning approaches
EVOLVER Optimization Evolves solutions using evolutionary code search

Mission Lifecycle

graph LR
    D[Discovery] --> R[Research]
    R --> P[Prototype]
    P --> E[Evolution]
    E --> S[Submission]

Each phase is implemented as a BaseNode in the state machine, returning either the next node or End[MissionResult].

Quick Example

import asyncio
from agent_k import LycurgusOrchestrator
from agent_k.core.models import MissionCriteria

async def main():
    async with LycurgusOrchestrator() as orchestrator:
        result = await orchestrator.execute_mission(
            competition_id="titanic",
            criteria=MissionCriteria(
                target_leaderboard_percentile=0.10,
                max_evolution_rounds=50,
            ),
        )
        print(f"Final rank: {result.final_rank}")

asyncio.run(main())

Getting Started

  • Installation

    Install AGENT-K with uv in minutes

    Install

  • Quick Start

    Run your first multi-agent mission

    Quick Start

  • Concepts

    Understand agents, toolsets, and the graph

    Concepts

  • API Reference

    Full API documentation

    API Reference

Technology Stack

Backend

Component Technology
Agent Framework Pydantic-AI
Orchestration Pydantic-Graph
Observability Pydantic Logfire
HTTP Client HTTPX

Frontend

Component Technology
Framework Next.js 16
Protocol AG-UI
Styling Tailwind CSS
Charts Recharts