Fat-Cat presents an LLM-native operating system framework designed to overcome critical challenges in current agent development: context management and control flow fragility. It targets developers building sophisticated LLM agents by offering a robust, debuggable, and evolving system that moves beyond traditional, state-heavy JSON-based approaches. Fat-Cat aims to make agent "thinking" transparent and persistent, enabling agents to learn and adapt over time.

How It Works

Fat-Cat re-imagines LLM agents by treating the LLM as the CPU, documents as memory (RAM), and tools as peripherals, with the framework acting as the kernel. It replaces fragmented JSON state management with Markdown documents as the global context, where each stage's output becomes a revision. This document-centric approach facilitates human readability and debuggability. The core operates through a four-stage metacognitive loop: Stage 1 analyzes intent and constraints, Stage 2 retrieves or dynamically learns strategies (triggering internet searches for new methodologies if needed), Stage 3 decomposes strategies into precise steps, and Stage 4 executes these steps. A Watcher Agent monitors execution for errors and deviations, providing runtime reflection and intervention capabilities.

Quick Start & Requirements

• Primary install / run command: Clone the repository (git clone https://github.com/your-repo/fat-cat.git), navigate into the directory, and run python scripts/install_full_pipeline_deps.py.
• Non-default prerequisites and dependencies: Python 3.10+, dependencies listed in requirements-full.txt. Configuration of LLM API keys is required in config/model_config.py. The framework is optimized for and recommends long-context models (32k+ context, e.g., Kimi-K2).
• Links: Repository URL: https://github.com/your-repo/fat-cat.git (placeholder).

Highlighted Details

• Document-Centric Context: Utilizes Markdown files for global state, enhancing transparency and debuggability over JSON.
• Metacognitive Loop: A hierarchical four-stage process (Analysis, Strategy, Decomposition, Execution) forces agents to "think twice" before acting.
• Dynamic Strategy Learning: Stage 2 agents can trigger "capability upgrades" via web searches (using Firecrawl/Tavily) to learn new problem-solving methodologies for novel issues.
• Runtime Monitoring: The Watcher Agent acts as a daemon to detect infinite loops, goal deviations, and allows for intervention or rollback.
• Benchmark Performance: Demonstrated significant improvements, including +12.58% on HotPotQA (multi-hop reasoning) and 95.3% accuracy on MBPP (Python code generation) compared to a React Agent baseline.

Maintenance & Community

No specific details regarding contributors, sponsorships, community channels (like Discord/Slack), or roadmaps were provided in the README.

Licensing & Compatibility

• License type: "[License Information]" - The license type is not specified in the provided README.
• Compatibility notes: Optimized for long-context LLMs (32k+ context recommended). Compatibility with commercial or closed-source linking is undetermined due to the unspecified license.

Limitations & Caveats

The framework's effectiveness is heavily dependent on the capabilities of the underlying LLM, particularly its long-context window support. The provided README uses a placeholder URL for the GitHub repository, and crucial licensing information is absent, posing potential adoption blockers.