infiAgent

infiAgent, also called MLA (Multi-Level Agent), is a Python agent framework from polyuiislab for long-running, complex tasks. Agents are defined through YAML configuration files rather than code, and the framework is distributed as an open-source GPL-3.0 repository, with a runtime installable from PyPI per the README.

What It Is

infiAgent lets users assemble general-purpose or semi-specialized agents, such as a research agent (Researcher) that handles literature search, experiments, figures, and LaTeX paper writing, or a computer-work assistant (OpenCowork) for tasks like organizing folders or building presentations. Agents run through a CLI, a Docker-based Web UI, a Python SDK, or a JSONL streaming mode intended for IDE plugins.

How the Architecture Works

The README describes a tree-structured hierarchy of single-purpose agents with minimal tool sets, where parent agents delegate to child agents serially. Rather than compressing conversation history, a thinking module periodically records the state of the file workspace and agents keep only recent actions. Long documents are queried through tools such as answer_from_pdf instead of being loaded into context, and outputs are saved as files. Task state is persisted so interrupted runs can be resumed.

Models, Skills and Memory

llm_config.yaml supports multiple providers per configuration, per-model API keys and base URLs, and separate models for execution, thinking, compression, and image understanding. Recent notes describe multi-deployment failover and atomic, crash-safe state writes. The framework follows the Agent Skills open standard, so skill folders can be discovered and loaded on demand. Memory is file-directory based: launching in the same workspace preserves history across sessions, with task history also indexed in a local SQLite database.

Setup Path

The README documents a Docker image that starts a Web UI on port 4242, a CLI mode, and a local pip installation requiring Python 3.9+ with Playwright Chromium. The Python SDK exposes run, add_message, fresh, reset_task, and task_snapshot calls for task-scoped control. The README warns that execute_command can run any command, and recommends Docker if a task may modify system files.