2026年9月18日 周五 · 开源推荐

Verified by executor via gh + full README fetch. Do not invent beyond this note.

1. Chosen project

• Repo: lsdefine/GenericAgent
• URL: https://github.com/lsdefine/GenericAgent
• Stars: ~14,216 ★ (forks ~1,654)
• License: MIT
• Language: Python
• Homepage/site: https://gaagent.ai (README also lists arXiv tech report https://arxiv.org/abs/2604.17091)
• Why worth trying today: Minimal (~3K-line seed) self-evolving computer-use agent — real browser/OS/ADB control + skill crystallization — a clear angle vs this week’s agent-harness flood (eve / hermes / deer-flow / headroom).

2. Problem it solves / who it’s for

Developers who want a local autonomous agent that actually drives the machine (browser with login sessions preserved, terminal/files, mouse/keyboard/vision, mobile via ADB), and that gets better over time by turning solved tasks into reusable Skills/SOPs — without adopting a heavy multi-service harness. Especially attractive if you care about token efficiency (README claims <30K context vs 200K–1M peers) and a try-it path that is clone + API key + TUI.

3. Key capabilities / usage (from FULL README)

Read full bilingual README (~803 lines) from gh api repos/lsdefine/GenericAgent/readme.

Design: “don’t preload skills, evolve them.” Core ≈ 3K LOC, ~100-line Agent Loop (agent_loop.py), 9 atomic tools.

Atomic tools: code_run, file_read, file_write, file_patch, web_scan, web_execute_js, ask_user, update_working_checkpoint, start_long_term_update.

Memory layers L0–L4: Meta Rules → Insight Index → Global Facts → Task Skills/SOPs → Session Archive.

Computer control: TMWebdriver = real Chrome/Chromium + extension (cookies/login preserved); claims strong bot-detection pass rates in README tables; computer-use / OCR / vision unlocked by *telling the agent* to set them up from SOPs in memory (not by reading long docs).

Frontends: TUI (frontends/tui_v3.py recommended EN), Streamlit (launch.pyw), Desktop portable, IM bots (Telegram/Discord/Lark; CN section also WeChat/QQ/WeCom/DingTalk).

Models: Claude / Gemini / Kimi / MiniMax etc. via mykey.py.

Self-evolution: new task → explore/install/debug → crystallize Skill → recall next time (WeChat read, HN digest, stock monitor, Gmail send examples).

Extras in README: Goal / Goal Hive / Morphling / Conductor modes; Datawhale tutorial; Skill Hub (Sophub); community GUIs.

4. Activity evidence

• Issues: Active user bug/feature traffic mid–late Sep 2026 (e.g. #809 open 2026-09-15 session-store mismatch; #808 “缺乏更新功能” open 2026-09-15 with contributor ZhulongNT reply 2026-09-17 CST-adjacent “git pull 即可”; #806/#804 model/UI bugs). Older closed threads show owner/contributor engagement (e.g. owner lsdefine on #672, 2026-07-14). Caveat: several open bugs lack immediate maintainer replies; ~170 open issues count via API (includes PRs in GitHub’s open_issues).
• Releases: Latest GitHub Release desktop-portable-v0.2.0 published 2026-08-24 (Windows/Linux/macOS portable Desktop; unsigned/ad-hoc signing caveats in notes). Project iterates mainly on main rather than frequent semver tags for the Python core.
• Commits: Last push 2026-09-14T03:43:01Z (~2026-09-14 11:43 Asia/Shanghai). Recent themes: constitution/prompt dedupe + remember-tool gates; Force Stop / urllib3 abort; vision backend docs; P2P hub reconnect backoff; TTFT/context window tuning; LLM channel switching. Open PRs still appearing through 2026-09-16 (#812 etc.).

5. Quick start (1–3 steps)

1. git clone https://github.com/lsdefine/GenericAgent.git && cd GenericAgent
2. uv venv && uv pip install -e ".[ui]" then cp mykey_template_en.py mykey.py and fill LLM API key

• Require Python 3.11 or 3.12 (README: do not use 3.14 / pywebview issues)

1. Launch: python frontends/tui_v3.py (TUI) or python launch.pyw (Streamlit)

• Optional one-liner installers in README (ga_install.sh / .ps1); then tell the agent e.g. *“Set up your web automation capability.”* / *“Probe this system and set up your computer-use capability.”*

6. Risks / caveats

• Full system control — runs code, drives browser/OS with user permissions; not a sandbox (similar risk class to coding agents).
• Last core commit ~4 days ago (2026-09-14) — still within 2 weeks, but quieter than omnigent/prime-agent’s daily merge firehose.
• Desktop release lag — portable Desktop tag is Aug 24; for latest behavior prefer git pull + Python path.
• Windows TUI flaky — README recommends Git Bash; some open desktop/session bugs (#806, #809).
• SOP language — preinstalled SOPs Chinese-first (agent can rewrite to English on request).
• Official channel disclaimer — only GitHub + gaagent.ai; DintalClaw sole authorized commercial partner.

7. Runners-up (memory shortlist)

1. omnigent-ai/omnigent (~10.0k★, Apache-2.0) — meta-harness for Claude Code / Codex / Cursor / Pi; v0.14.0 @ 2026-09-15; commits through 2026-09-18; very strong activity but overlaps this week’s harness theme → keep as next-day candidate if wanting orchestration UX.
2. withastro/flue (~8.3k★, Apache-2.0) — “sandbox agent framework”, TypeScript harness; @flue/cli@2.0.8 @ 2026-09-16; commits through 2026-09-17. Also strong but harness-adjacent: PrimeIntellect-ai/prime-agent (~21.0k★, MIT, self-improving RLM; v0.9.5 @ 2026-09-16; hyper-active Sep 17 commits) — note vouched-only issue filing (community reports auto-closed). Fresh non-harness watch: Tencent/BrowserSkill (~4.3k★, MIT, CLI+extension for real logged-in browser; cli-v0.3.0 @ 2026-09-17).

Still avoid (already recommended): DeusData/codebase-memory-mcp, volcengine/OpenViking, vshulcz/deja-vu, pizza-bot-app/pizza-bot, HKUDS/nanobot, bytedance/deer-flow, vercel/eve, headroomlabs-ai/headroom, NousResearch/hermes-agent.

8. Verification checklist

• Full README read (English + Chinese sections via API base64 decode, 803 lines)
• Issues sampled (open Sep list + closed history + issue #808 view; contributor reply 2026-09-17)
• Releases checked (gh release list + gh release view desktop-portable-v0.2.0)
• Commits checked (gh api .../commits?per_page=8, last 2026-09-14; themes noted)
• Shortlist peers refreshed (omnigent, flue, prime-agent) + light expand search; pick chosen for distinct computer-use + self-evolution angle