SuperIntelligence: Why the Future of AI is a File System (CORAL)

2026-04-07 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

The Coral project, developed by institutions including MIT, National University of Singapore, Stanford, and Meta, introduces an "Advanced Intelligence" (ADI) infrastructure for self-evolving, multi-agent AI systems. Instead of enhancing the intelligence of Large Language Models (LLMs) directly through reinforcement learning, Coral focuses on optimizing the external file system and environment around frozen LLMs. This autonomous infrastructure, designed for open-ended complexities, utilizes a standardized hierarchical file system where each AI agent operates within a local Git work tree. Agents communicate and coordinate implicitly by reading and writing to a shared persistent memory, specifically a hidden `.coral` public directory containing "attempts" (function evaluations), "notes" (textual hypotheses), and "skills" (executable code modules). The system employs a heartbeat intervention protocol to prevent agents from getting trapped in local optima, forcing them to externalize intermediate findings or attempt orthogonal approaches. This gradient-free search operates at inference time, with learning entirely contextualized through in-context memory accumulation within the file system.

Key takeaway

For research scientists exploring autonomous agent systems, Coral presents a novel approach to scaling intelligence without modifying core LLMs. You should consider this file-system-centric methodology for its potential to manage open-ended tasks and multi-agent coordination, especially if direct LLM fine-tuning is cost-prohibitive or technically challenging. Be mindful of the significant computational costs associated with multi-agent runs, and implement thresholds to manage expenses.

Key insights

Coral enhances AI system intelligence by optimizing external file systems and multi-agent coordination, not by training LLMs.

Principles

• Externalize AI learning to file systems.
• Utilize multi-agent, autonomous evolution.
• Employ gradient-free search at inference.

Method

Coral uses a shared, hierarchical file system with Git work trees for agents. It stores attempts, notes, and skills, and employs a heartbeat protocol to trigger externalization or orthogonal approaches, all while LLMs remain frozen.

In practice

• Implement Git work trees for agent isolation.
• Structure shared memory with attempts, notes, skills.
• Use heartbeat protocols to escape local optima.

Topics

• Coral Infrastructure
• File System Optimization
• Multi-Agent Systems
• LLM Harness
• Gradient-Free Search

Best for: Research Scientist, AI Scientist, AI Engineer, AI Architect

Related on AIssential

• 🥇Top AI Papers of the Week — AI advancements focus on autonomous agentic systems, efficient LLM inference, and specialized foundation models.
• DR-Venus: Towards Frontier Edge-Scale Deep Research Agents with Only 10K Open Data — Small models can achieve strong deep research capabilities through high-quality data and effective utilization.
• EVE: A Domain-Specific LLM Framework for Earth Intelligence — EVE provides an open, end-to-end framework for domain-specialized LLMs in Earth Intelligence, outperforming general models.
• Cloud Ant Colonies — Complex AI behavior emerges from simple agents coordinating indirectly through a shared environment.
• A Hierarchical Multi-Agent System for Autonomous Discovery in Geoscientific Data Archives — A hierarchical multi-agent system autonomously discovers and analyzes geoscientific data, improving reuse and reducing manual effort.
• Yann LeCun Pushes Back on AGI Hype at India AI Summit: "Superintelligence Isn't Close" — True AI requires a "world model" for physical interaction, a capability current LLMs and systems lack.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.