GenericAgent

What It Is

GenericAgent is a minimalist, self-evolving autonomous Agent framework with a core of only about 3K lines of code. Through 9 atomic tools plus an Agent Loop of about 100 lines, it gives any LLM system-level control over a local computer, covering browsers, terminals, file systems, keyboard and mouse input, screen vision, and mobile devices (ADB).

The design philosophy is "do not preset skills; gain capabilities through evolution." Each time it solves a new task, GenericAgent automatically solidifies the execution path into a Skill for direct invocation later; the longer it is used, the more skills accumulate, forming a personal skill tree grown from 3K lines of seed code. The official self-bootstrapping proof is: from installing Git and running git init to every commit message, this repository was completed autonomously by GenericAgent, and the author never opened a terminal even once throughout the process.

Main Capabilities

• System-level control: Interacts with the outside world through 9 atomic tools, covering browsers, terminals, file systems, keyboard and mouse input, screen vision, and mobile devices (ADB).
• Autonomous execution loop: Perceive environment state -> task reasoning -> invoke tools for execution -> write experience into memory -> loop. The core loop is about a hundred lines of code (agent_loop.py).
• Layered memory system: Memory continuously accumulates during task execution, enabling the Agent to gradually form a stable and efficient way of working.
• Dynamic capability expansion: Through code_run, it can dynamically install Python packages, write new scripts, call external APIs, or control hardware at runtime, and solidify temporary capabilities into permanent tools.
• Skill self-evolution: Task execution paths automatically crystallize into reusable Skills, forming a dedicated skill tree through long-term use.
• Multiple frontends: Provides TUI (v2/v3), desktop GUI, as well as IM frontends such as Telegram, Discord, WeChat, QQ, Feishu/Lark, WeCom, and DingTalk.

Installation and Usage

⚠️ Python version: Python 3.11 or 3.12 is recommended. Do not use Python 3.14, as it is incompatible with pywebview and some dependencies.

One-click installation automatically prepares an independent Python environment, Git, project files, and the desktop client without polluting the system environment. Startup methodFormat:

• Windows: Double-click frontends/GenericAgent.exe
• Linux / macOS: Run python launch.pyw in the installation directory

Officially, it is more recommended for the Agent to bootstrap the environment during use, rather than manually installing all dependencies in advance. In addition, the installation guide (installation.md / installation_zh.md) and installation scripts are located in the repository's assets/ directory, and can be read before execution.

Supported interfaces include: desktop GUI, TUI v3 (frontends/tui_v3.py) and TUI v2 (frontends/tuiapp_v2.py), as well as IM frontends (Telegram, Discord, WeChat, QQ, Feishu/Lark, WeCom, DingTalk). For detailed configuration of IM frontends, the official documentation says you can ask GenericAgent itself directly.

About TUI: based on Textual or prompt_toolkit + rich. If rendering is abnormal, the official recommendation is to first run pip install -U textual or pip install -U prompt_toolkit rich; on Windows, Git Bash is recommended rather than PowerShell/cmd, and you can also let GA fix compatibility issues itself (the official documentation provides the corresponding prompt template).

Unlocking Advanced Capabilities

Official note: advanced capabilities are unlocked by "instructing the Agent" rather than by reading documentation or installing extensions. Each instruction causes GA to read its preinstalled SOPs (practical manuals in memory), install missing parts, adapt to the user's operating system, and persist the results into its own memory. The preinstalled SOPs are in Chinese, and GA can read them natively; if an English knowledge base is needed, you can instruct it to rewrite the preinstalled SOPs into English (code, paths, and error strings remain unchanged). The SOPs were refined on Windows, and cross-platform adaptation on macOS/Linux (window enumeration, input control, screenshots, etc.) is completed by GA itself.

Published Modes and Updates (Summary of Official Roadmap & News)

• 2026-05-23 — TUI v3 released (frontends/tui_v3.py): block-based scrollback review, resize reflow, independent color scheme per terminal, aligned with v2 features.
• 2026-05-18 — Morphling mode: project-level capability devouring; after extracting goals and test cases from any external repository, decide separately for each core component whether to call, rewrite, or discard it (memory/morphling_sop.md).
• 2026-05-17 — Goal Hive mode: a multi-worker collaborative version of Goal, where Master/Worker coordinate through BBS to advance long-horizon goals (memory/goal_hive_sop.md).
• 2026-05-15 — Desktop GUI released: one-click installation includes a runnable desktop client (frontends/GenericAgent.exe); developers can start it with python launch.pyw.
• 2026-05-14 — Conductor sub-Agent orchestration: dispatch, supervise, and automatically clean up parallel sub-Agents, complementing the /btw side-path sub-Agent.
• 2026-05-12 — TUI v2 released (frontends/tuiapp_v2.py): supports image paste folding, file paste, block deletion, Ctrl+C copy, history navigation, and /llm /export /continue selectors.
• 2026-05-08 — Goal mode (reflect/goal_mode.py): a time-budget-driven self-propelled loop that does not deliver early before the budget is reached.
• 2026-04-21 — Technical report published to arXiv: GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization.
• 2026-04-11 — Introduced L4 session archive memory and integrated scheduler cron scheduling.
• 2026-03-23 — Added support for personal WeChat access as a Bot frontend.
• 2026-03-10 — Released a million-scale Skill library.
• 2026-03-08 — Released "Government Lobster" Dintal Claw centered on GenericAgent.
• 2026-03-01 — Reported by Machine Heart.
• 2026-01-16 — GenericAgent V1.0 public release.

Evaluation and Data

The official statement says GenericAgent is evaluated across five major dimensions, with baselines including Claude Code, OpenAI CodeX, and OpenClaw, evaluated respectively on Claude Sonnet 4.6, Claude Opus 4.6, GPT-5.4, and MiniMax M2.7 foundations. The complete evaluation dataset and results are available at JinyiHan99/GA-Technical-Report.

Regarding browser authenticity: the GA Web tool is driven by TMWebdriver (a local WebSocket service plus a Chrome extension), running in a real, persistent Chrome/Chromium session rather than a one-off headless sandbox, so it can retain cookies, login state, extensions, GPU/WebGL behavior, and a normal browser session fingerprint. The official statement says that for reCAPTCHA v3, 0.9 is a high-confidence human-likeness score returned by the risk-control model, usually sufficient to pass common production environment thresholds.

Applicable Scenarios

• Automation tasks that require letting an LLM directly control a local computer (browser, terminal, file system, keyboard and mouse, screen).
• Long-term scenarios where the Agent is expected to continuously accumulate personalized skills and working patterns during use.
• Projects that are sensitive to token consumption and hope to gain system control capability with a smaller code core.
• Scenarios that remotely drive the Agent through IM frontends (Telegram, Discord, WeChat, QQ, Feishu/Lark, WeCom, DingTalk).

Limitations and Notes

• The Python version must be 3.11 or 3.12; 3.14 cannot be used.
• The official statement shows: the official channels are only this GitHub repository and https://gaagent.ai；DintalClaw is currently the only officially authorized business partner; other third-party websites, institutions, organizations, or individuals using the GenericAgent name are all unofficial unless explicitly listed.
• License: MIT License, see the repository LICENSE for details.
• Specific fees, prices, and commercial terms are not yet provided.