Active-ARC

A Python SDK for building agents that solve Active-ARC tasks.

Installation

pip install active-arc

Or with uv:

uv add active-arc

Quick Start

from active_arc import Agent, Swarm, SessionState, Grid

class MyAgent(Agent):
    MAX_QUERIES = 50

    def query(self, session: SessionState) -> Grid | None:
        """Choose an input grid to query the oracle. Return None to stop querying."""
        # Your logic here
        return [[1, 2], [3, 4]]

    def submit(self, session: SessionState) -> Grid:
        """Predict the output grid for session.test_input."""
        # Your logic here
        return session.test_input

# Run on tasks
swarm = Swarm(
    tasks=["007bbfb7"],
    agent_class=MyAgent,
)
swarm.run()

Configuration

Set your API key as an environment variable:

export ACTIVE_ARC_API_KEY="your_api_key_here"

Or create a .env file:

ACTIVE_ARC_API_KEY=your_api_key_here

Agent API

Agents must implement two methods:

Session State

The SessionState object provides:

• session.task_id - Current task ID
• session.query_count - Number of queries made
• session.queries - List of QueryRecord objects (input/output pairs)
• session.test_input - The test input grid (available in test phase)
• session.phase - Current phase (learning, test, or completed)

Optional Methods

Override these for custom reasoning metadata:

def get_query_reasoning(self, session: SessionState, query_input: Grid) -> Any:
    """Return metadata to store with this query."""
    return {"model": "gpt-4o", "tokens": 150}

def get_submit_reasoning(self, session: SessionState, answer: Grid) -> Any:
    """Return metadata to store with the submission."""
    return {"confidence": 0.85}

Swarm API

The Swarm class runs agents on multiple tasks:

from active_arc import Swarm, Random

# Using a built-in agent
swarm = Swarm(tasks=["007bbfb7", "00d62c1b"], agent="random")

# Using a custom agent class
swarm = Swarm(tasks=["007bbfb7"], agent_class=MyAgent)

# With explicit configuration
swarm = Swarm(
    tasks=["007bbfb7"],
    root_url="https://api.active-arc.com",
    agent_class=MyAgent,
    tags=["experiment-1", "v1.0"],
)

scorecard = swarm.run()
print(f"Solved: {scorecard.solved}/{scorecard.attempted}")

Built-in Agents

Random Agent

Makes random queries and returns the test input as the answer:

from active_arc import Swarm

swarm = Swarm(tasks=["007bbfb7"], agent="random")
swarm.run()

LLM Agent

Uses OpenAI's API to make intelligent queries:

from active_arc import Swarm

# Requires OPENAI_API_KEY environment variable
swarm = Swarm(tasks=["007bbfb7"], agent="llm")
swarm.run()

CLI Usage

Run agents from the command line:

# Run random agent on a specific task
python main.py -a random -t 007bbfb7

# Run with custom tags
python main.py -a random -t 007bbfb7 --tags "experiment,v1.0"

Data Types

from active_arc import (
    Grid,          # list[list[int]] - 2D grid of integers 0-9
    SessionState,  # Current session state
    SessionPhase,  # Enum: LEARNING, TEST, COMPLETED
    Scorecard,     # Results for all tasks
    TaskCard,      # Results for a single task
    QueryRecord,   # Input/output pair from a query
)

License

MIT License