Summary

ARCgentica is an agentic AI system designed to solve the Abstract Reasoning Corpus (ARC) challenges using Large Language Models (LLMs). It targets researchers and AI developers aiming to push the boundaries of abstract reasoning and AGI benchmarks, offering a framework that achieves high performance on the ARC-AGI-2 public evaluation.

How It Works

The system employs a multi-agent architecture where sub-agents analyze input-output grid examples, programmatically generate potential solutions, and then evaluate these programs against test inputs. This decomposition allows for a systematic approach to complex reasoning tasks, leveraging LLMs for code generation and logical deduction.

Quick Start & Requirements

• Primary Install/Run: Clone repositories (arcgentica, agentica-server), install server dependencies (uv sync && uv pip install numpy scikit-image scipy sympy), start the server (uv run src/application/main.py --inference-token=<your-key> --inference-endpoint <endpoint> --sandbox-mode='no_sandbox' --max-concurrent-invocations 1200 --port 2345), set environment variable (export S_M_BASE_URL=http://localhost:2345), and run the main script (uv run python main.py --model anthropic/claude-opus-4-6).
• Prerequisites: Python 3.12.11, uv package manager, and an API key for a supported model provider (OpenAI, Anthropic, or OpenRouter).
• Resource Footprint: Requires LLM API access and local server setup.
• Docs/Demos: Blog Post, ARC Prize.

Highlighted Details

• Achieved 85.28% on the ARC-AGI-2 Public Evaluation using Anthropic's Claude Opus 4.6.
• Reported cost per task is approximately $6.94.
• Average of 2.6 agents used per problem.
• Mean time per task is around 1511.9 seconds.

Maintenance & Community

• Developed by Symbolica AI.
• Community channels include Discord and Twitter.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive MIT license allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The system relies heavily on external LLM API performance and availability. The setup requires running a local server and configuring API keys, which may introduce complexity. The --sandbox-mode='no_sandbox' setting for the server warrants careful consideration regarding execution environment security.