GenericAgent — a self-evolving agent grown from a 3.3K-line seed

9 + 100

Nine tools and a hundred-line agent loop. That's the entire core of GenericAgent. Layer an LLM's coding ability on top and the agent operates a browser, terminal, file system, keyboard, mouse, screen vision, and ADB-controlled mobile devices at the system level. 8.8K stars, +320 in 24 hours, with an arXiv companion paper (2604.20710).

The pitch: start from a minimal seed, grow a full LLM-driven desktop control surface.

Why minimalism

Frameworks like SuperAGI and AutoGPT bundle large tool catalogs upfront. Context bloat and tool-selection errors scale with that catalog.

GenericAgent flips it. "Atomic tools cap at nine; new capabilities emerge as the LLM writes code." Solved tasks crystallize into "skills" stored in a skill tree. Repeat tasks reuse skills, cutting token cost roughly 6x — the paper's headline result.

[IMG#1]

Six features

Feature	Description
9 atomic tools	Browser, terminal, file, keyboard, mouse, vision, ADB, interface, self-modify
~100-line agent loop	Minimal core, readability-first
Self-evolving skill tree	Solved tasks become permanent tools
Layered memory	30K context preserved, 6x token savings reported
Dynamic runtime install	pip packages, external APIs, hardware on demand
5 frontends	Streamlit, QQ, Telegram, Feishu, WeCom, DingTalk

Where most agents grow tool catalogs, GenericAgent goes the opposite way — keep tools small, grow skills.

Tech stack

• Language: Python
• UI: Streamlit (desktop), 5 messenger bots
• External integrations: ADB (mobile), Selenium/Playwright (browser)
• LLM: OpenAI / Anthropic API

Three layers — (1) Agent Loop (100 lines), (2) Atomic Tools (9 modules, ~300 lines each), (3) Layered Memory (context manager). New skills accumulate as code-plus-example pairs in a "Persistent Skills" region.

Repo comparison

Repo	Stars	License	Position
lsdefine/GenericAgent	8.8K	Apache-2.0	Minimal seed + self-evolving skill tree
TransformerOptimus/SuperAGI	23K	MIT	Full tool catalog
Significant-Gravitas/AutoGPT	168K	MIT	Early autonomous agent, full toolset
e2b-dev/awesome-ai-agents	10K	MIT	Curation list (not a runtime)

GenericAgent is fourth on stars, but its self-evolving design clearly differs from peers. The design point matters more than the absolute star count here.

[IMG#2]

Why now — ecosystem context

Three trends converged. (1) Computer Use and OSWorld benchmarks made "screen-driven agents" a measurable category. (2) Skill-tree paradigms à la Voyager (originally MineDojo) are migrating into general LLM agents. (3) An arXiv companion paper pulled in academic users immediately.

Getting started

git clone https://github.com/lsdefine/GenericAgent
cd GenericAgent
cp mykey_template.py mykey.py   # add API keys
python launch.pyw

Common gotchas — mykey.py must contain OpenAI/Anthropic keys or it halts. macOS ADB requires brew install android-platform-tools.

Limits and outlook

Limits — (1) Security: LLM has code execution rights with weak isolation; risky in production environments. Docker-isolation PR is in flight. (2) Skill-tree retrieval cost rises with size; the 6x token-savings claim should be re-measured beyond ~1,000 skills.

Outlook — next six months likely brings (a) Docker/gVisor isolation, (b) a skill marketplace for community sharing, (c) MCP integration to expose skills to other agents.

[IMG#3]

3-Line Summary

• 9 atomic tools + 100-line agent loop deliver system-level control; 8.8K stars.
• Solved tasks become permanent skills; reported 6x token savings via the skill tree.
• arXiv 2604.20710 companion paper grounds the "small tools, growing skills" inversion.

References

• GitHub — lsdefine/GenericAgent
• arXiv — 2604.20710
• SuperAGI — competing full-toolset framework
• Voyager — skill-tree origin
• Hacker News — GenericAgent discussion