I've gathered and cross-checked the live data. Here's my report. ## What this list actually is The page you pointed me at — [30 Best AI Agents GitHub Repos (September 2026)](https://findarepo.com/categories/ai-agents) — is a **momentum leaderboard**, not a stars leaderboard. It says so itself ("ranked by star velocity findarepo measures itself … no paid placement"), and the [published methodology](https://findarepo.com/about/methodology/) confirms the math: `score = log1p(velocity)×32 + min(25, relative_pct×2.5) + log10(stars+1)×5 + activity×10`, built from **daily star snapshots taken through the GitHub REST search API** (GitHub publishes no star history, so the site constructs one). Data-refresh date on every page: **Sep 12, 2026**. ## What changed **1. The leaderboard has rotated from frameworks to "harnesses," skills and plumbing.** The current top 5 is ponytail (136k ★, +9.2k/7d), deepseek-harness (221k ★, +8.4k), ECC (257k ★, +7.8k), orca (67k ★, +5.1k) and OmniRoute (65k ★, +3.6k) ([category page](https://findarepo.com/categories/ai-agents)). I verified every headline star count directly against GitHub's API — e.g. [affaan-m/ECC = 256,899 ★ / 38,444 forks, created 2026-01-18](https://api.github.com/repos/affaan-m/ECC), [deepseek-ai/deepseek-harness = 221,420 ★, created 2026-08-13](https://api.github.com/repos/deepseek-ai/deepseek-harness), [DietrichGebert/ponytail = 136,410 ★, created 2026-06-12](https://api.github.com/repos/DietrichGebert/ponytail), [NousResearch/hermes-agent = 244,822 ★](https://api.github.com/repos/NousResearch/hermes-agent), [stablyai/orca = 67,171 ★](https://api.github.com/repos/stablyai/orca). Spot-checking 15 more mid-list repos via `api.github.com/search/repositories` matched the page's rounded totals (firecrawl 179,468; graphify 117,164; claude-mem 93,725; deer-flow 82,323; ruflo 72,215; mem0 65,183). **The star totals are real; the "+x/7d" deltas are the site's own measurements and can't be independently confirmed.** **2. The old guard has fallen off this list.** An April 2026 top-20 listicle was led by [AutoGPT (183k), Langflow (146k), Dify (136k), LangChain (132k), Gemini CLI (100k), browser-use (86k)](https://fungies.io/top-github-repositories-ai-agent-frameworks-2026/). Today AutoGPT is at 187,275 ★ and Langflow at 154,651 ★ — active, but **absent from findarepo's 30** because they're not moving fast enough; LangChain survives only at #24 (+501/7d). Microsoft's AutoGen, a 2025 staple, shows a last push of **2026-04-15** — roughly five months idle. A DataAIHub snapshot from [Aug 14, 2026](https://www.dataaihub.co/research/top-ai-agent-github-repositories-2026) had hermes-agent #1 among 63 agent-classified repos at 230.2k ★; it's at 244.8k now (+~14.6k in four weeks), while LangChain went 144.2k → 146.2k — a 7× velocity gap that explains the churn. **3. A model lab entered the harness business.** DeepSeek open-sourced **Harness v0.1** on **Aug 13, 2026** under MIT: "Agent = Model + Harness," every capability (models, tools, skills, sessions, sandboxes, storage, loop, scheduling, UI) a swappable plugin on the **Cordis kernel** ([official announcement](https://www.deepseek.com/harness/en/)). The repo already has 221k ★ and has **issues and PRs disabled** per the API payload. The ecosystem is forming around it fast — findarepo's [trending page](https://findarepo.com/trending/) lists dsh-desktop (26k ★) and awesome-dsh-plugin (15k ★) under "New & rising." **4. "Skills" packages now outrank agent frameworks.** ponytail is a Claude Code/Codex plugin that makes agents write less code ([README: –54% LOC, –20% cost, –27% time on a headless Claude Code benchmark](https://github.com/DietrichGebert/ponytail)); ECC is a single-maintainer harness-optimization pack shipping 68 agents, 292 skills and adapters for 7+ harnesses ([repo](https://github.com/affaan-m/ECC)). GitHub's own [weekly trending page](https://github.com/trending?since=weekly) — an independent source — is dominated by the same cluster: i-have-adhd, ECC, ponytail, context-mode, gods-eye-view, archify, diagram-design, humanizer, hyperframes, plus **openai/skills (26,987 ★)** and **openai/plugins (6,510 ★)**. Anthropic's official plugin directory sits at 36,168 ★. **5. Day-over-day churn is extreme.** Comparing the [Sep 11 digest](https://findarepo.com/digest/2026-09-11/) with the [Sep 12 digest](https://findarepo.com/digest/): i-have-adhd went +13k → +16k (43k ★) and is #1 on GitHub's weekly trending too; gods-eye-view and google/artemis entered; Edge0 and VoiceStudio dropped out entirely. ## Why it matters - **Ranking by velocity means this list measures attention, not dependability.** LangChain (146k ★) sits at #24, AutoGPT (187k ★) doesn't appear at all. Picking a production stack off this page is picking whatever is trending this week — exactly the failure mode the page's own "How to choose" note warns about ("an agent framework that hasn't merged a PR in two months is already behind the model releases it wraps"). - **The harness is becoming the contested layer.** DeepSeek ships one; OpenAI ships skills/plugin catalogs; Anthropic ships an official plugin directory; YC-backed Orca ([repo, "fleet of parallel agents"](https://api.github.com/repos/stablyai/orca)) and Rust runtimes like herdr sit on top. DeepSeek explicitly makes no plugin privileged, "including the DeepSeek model provider itself" — a portability bet that, if it wins, makes today's harness lock-in moot. - **Memory, context and routing are now first-class categories**: claude-mem (93.7k ★), mem0 (65.2k ★), supermemory (29.6k ★), headroom (71.7k ★, token compression), context-mode (22.3k ★), and free gateways OmniRoute (65k ★) / 9router (28.5k ★) / CLIProxyAPI (51.5k ★). - **Open-core is the business model**: ECC is MIT-forever but monetized via a $19/seat/mo GitHub App for private repos — "Sponsors and Pro subscribers fund the work" ([repo](https://github.com/affaan-m/ECC)). ## Caveats worth flagging - The category page's **"formula on the methodology page" link is broken** — `findarepo.com/methodology` 404s; the real page is `/about/methodology/`. - **The category is noisy.** It includes non-agent repos: firecrawl (scraping API), ragflow (RAG engine), pascalorg/editor (3D architectural editor), asgeirtj/system_prompts_leaks, ai-job-search and book-to-skill. DataAIHub makes the same admission about its own classifier (it lumps MLflow, Nx, ToolJet and "a curated app list" into "agents"). - **Star inflation is a live risk**, and findarepo says so: it cites independent research on millions of fake stars and ships a [star-credibility checker](https://findarepo.com/check/) scoring fork ratio, issue activity and stargazer-account quality. It calls its output "a signal, not a verdict." - **Other trackers are less reliable right now.** [OSSInsight's trending AI page](https://ossinsight.io/trending/ai) rendered **zero repos** and warns that GitHub's public events feed "has been incomplete since 2026-03-01," so its growth figures are "lower bounds." - **Verify DeepSeek tooling claims from primary docs.** A [SitePoint tutorial](https://www.sitepoint.com/deepseek-harness-developer-preview/) describes Harness as Python 3.10–3.12 with a `harness.config.yaml`, while DeepSeek's official page and the GitHub API show a **TypeScript/Node** tool launched via `npx @deepseek-ai/dsh web`. Third-party guides are already drifting from the source. **Strongest sources:** [findarepo AI Agents ranking](https://findarepo.com/categories/ai-agents) · [findarepo methodology](https://findarepo.com/about/methodology/) · [DeepSeek's official Harness announcement](https://www.deepseek.com/harness/en/) · [GitHub weekly trending](https://github.com/trending?since=weekly) · [GitHub API entries for the top repos](https://api.github.com/repos/affaan-m/ECC) · [DataAIHub Aug 2026 snapshot](https://www.dataaihub.co/research/top-ai-agent-github-repositories-2026) · [findarepo daily digests](https://findarepo.com/digest/2026-09-12/).