This framework provides an unlimited runtime AI agent system designed for complex, long-running tasks without context degradation. It enables users to build domain-specific, state-of-the-art agents, particularly for research and scientific computing, by leveraging configuration files and a novel multi-level agent architecture. The primary benefit is the ability to handle extensive workflows, such as academic paper writing and scientific simulations, autonomously and persistently.

How It Works

InfiAgent employs a multi-level agent hierarchy, orchestrating agents in a tree structure for focused roles and clear delegation. Its core innovation lies in a file-centric architecture and a "Ten-Step Strategy" where agent state is updated based on file system changes every ten steps, eliminating the need for context compression. A nested attention mechanism extracts only relevant information from large documents, preserving context efficiency. Batch file operations and a task ID system tied to workspace paths ensure persistent memory across sessions, allowing for truly unlimited runtime without performance degradation.

Quick Start & Requirements

The recommended installation is via Docker.

1. Install Docker.
2. Pull the image: docker pull chenglinhku/mlav3:latest
3. Run Web UI mode:

   docker run -d --name mla \
     -e HOST_PWD=$(pwd) \
     -v $(pwd):/workspace$(pwd) \
     -v ~/.mla_v3:/root/mla_v3 \
     -v mla-config:/mla_config \
     -p 8002:8002 -p 9641:9641 -p 4242:4242 -p 5002:5002 \
     chenglinhku/mlav3:latest webui
   

   Access at http://localhost:4242.
4. Configure API keys at http://localhost:9641 by editing run_env_config/llm_config.yaml.

Local installation requires Python >= 3.10, pip install -e ., and playwright install chromium.

Highlighted Details

• Unlimited Runtime: Achieved via file-centric state management and a ten-step strategy, avoiding context degradation.
• End-to-End Research Workflows: Capable of literature search, experiments, plotting, and LaTeX paper writing, with claims of producing papers that pass EI/IEEE conference peer reviews.
• Zero Context Compression: State is managed via the file system, eliminating the need for context compression techniques.
• Multi-Level Agent Hierarchy: Serial execution with tree-structured orchestration for focused agents and clear delegation.
• Batch File Operations: Uses list-based parameters for tool calls (e.g., file_read(paths=[...])) to save tokens.

Maintenance & Community

The project shows active development with frequent updates in early 2026 addressing bugs and adding features like Web UI enhancements and improved LLM integration. Key contributors include Yu, Chenglin and Wang, Yuchen. Contact is available via email (yuchenglin96@qq.com, etc.) and GitHub.

Licensing & Compatibility

The specific open-source license is detailed in a LICENSE file, but not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not detailed.

Limitations & Caveats

The framework currently supports Python projects for coding tasks, with potential for other languages in the future. A temporary fix for the Web UI is noted, with a full resolution pending. Older versions had restricted command execution; newer versions allow all commands, but using Docker is recommended for tasks that might modify system files.