Every six hours, my little open-source bot snapshots GitHub Trending into a clean digest so I can watch where the developer world is going without doom-scrolling. I've been doing this for months, and this week's snapshot is one of the clearest signals I've seen: the trending page has stopped being about models and started being about how agents are organized, equipped, and verified.
Here is the live data — repos created in the last 7 days, sorted by stars:
| Repo | Stars | What it is |
|---|---|---|
XiaoDuoYa/codex-with-chatgpt |
~2.3k | ChatGPT as the planning brain, Codex as the hands — a two-agent harness |
Nanako0129/sepia |
~1.6k | A "De-AI" writing skill for Agent Skills-compatible agents, 77+ rules |
cbrock84/headcount |
~1.1k | Claude Code structured as a company: 15+ departments, 125+ skills |
MetaMask-AI/metamask-desktop |
~1.2k | The MetaMask desktop app (Ethereum browsing) |
anthropics/commerce-agents |
~570 | Anthropic's official blueprint for shopping/merchant agents |
2akouwu/reverify |
~600 | "Verified reverse engineering": AI grounded in deterministic tools |
Player-YN/PawWork_ZhuaZhua |
~560 | A selection-first web agent for Chrome |
subsy/skill-cabinet |
~370 | Tooling for managing skills themselves |
Five of the top eight are agent-related. Last month, the same slot would have been filled by model repos. That is a genuine category shift, and it clusters into three patterns worth stealing:
1. SKILL.md is becoming a distribution format
sepia, headcount, and skill-cabinet all treat a skill — a folder with a SKILL.md plus scripts and references — as the atomic unit of software. Not a package, not a container: a skill. Why it matters: skills are loaded on demand into an agent's context, so distribution cost drops toward zero compared to libraries that must be installed and imported. The ecosystem that already emerged around Agent Skills (CLI installers, marketplaces, registries) looks a lot like npm in 2011.
2. Agent brains are splitting — planner vs. executor
codex-with-chatgpt (ChatGPT thinks, Codex works) and anthropics/commerce-agents both encode the same architectural bet: don't make one agent do everything. Give one model the reasoning/planning role and another the execution harness. This matches what I've learned running automated publishing pipelines for months — a cron agent that both decides and executes fails silently; separating the planner from the executor and giving the executor external checks is what actually catches garbage before it ships.
3. Verification is the new moat
reverify (AI reverse engineering checked against deterministic tools) is the sharpest example: when an agent produces output, the winning pattern is not better prompts but an external verifier the agent can't argue with. In my own stock-analysis and content pipelines, every time I replaced "trust the agent's success report" with an independent check (e.g. live market data, or a hard gate on output quality), reliability jumped. The models change; the need for falsifiable checks does not.
What to actually do this week
- If you use Claude Code / Codex / Cursor: try packaging one of your repetitive workflows as a skill. A skill you can install on any agent beats a script you have to re-explain.
- If you build agents: budget 30% of your effort for a verifier layer, not for prompt polish. That is the difference between a demo and a tool.
- If you watch trends: don't trust your memory — snapshot the data. My aggregator (zero-cost, zero-API, updates every 6h) is free to use: https://felixwang007.github.io/awesome-content-tools/ (repo:
Felixwang007/awesome-content-tools).
One more plug that fits the theme: the skill economy is not just an English-language thing. I publish Chinese-language skills for AI agents too — including an A-Share stock analysis expert (three-pillar system: technical + fundamental + sentiment, 24 indicators) on the xiaping.coze.com skill marketplace, with the free open-source version in the same GitHub. If you trade Chinese markets or just want to see a real-world skill in action, both are linked from the repo above. The skills are the product now — the repos are just the packaging.
Top comments (0)