GitHub Is Packed With Free Agent Teams—Of the 135, You Probably Only Need 3

GitHub has recently gained a batch of free Claude Code team packs. The starting point was Anthropic turning Agent Skills into an open specification, after which the community packaged agents, skills, commands, and hooks into installable marketplaces. The featured tweet lists 10 repos; one pack contains 135 agents, 35 skills, 42 commands, 176+ plugins, and 20 hooks, while another directly provides 100+ specialized subagents—all free.

First, here are the 10 repos:

01 wshobson/agents, a plugin marketplace; the same set of agents can be installed into Claude Code, Codex, Cursor, and Copilot, so your configuration carries over when you switch tools.

02 VoltAgent/awesome-claude-code-subagents, 100+ specialized subagents, covering tasks most engineering weeks will encounter.

03 rohitg00/awesome-claude-code-toolkit, 135 agents, 35 curated skills, 42 commands, 176+ plugins, 20 hooks, and 14 MCP configurations.

04 anthropics/skills, the official Agent Skills repo. Before copying anything, look at what a working SKILL.md looks like.

05 netresearch/claude-code-marketplace, skills organized according to open standards, usable across Claude Code, Cursor, Copilot, Codex, Gemini CLI, and 30+ other agents.

06 disler/claude-code-hooks-mastery, a hooks reference, with an emphasis on making agents stop according to your rules and interrupt you less.

07 karanb192/claude-code-hooks, an installable hooks marketplace, categorized into security, cost, observability, and productivity.

08 affaan-m/claude-swarm, multi-agent orchestration: decomposes tasks, coordinates agents, and shows the full run in a terminal UI.

09 am-will/swarms, turns orchestration into skills, loading coordination logic only when a task actually needs it.

10 eugeniughelbur/obsidian-second-brain, uses ordinary markdown in an Obsidian vault for persistent memory, shared across Claude Code, Codex, and other tools.

Behind the hype are two problems. The first is the cost of choosing. 135 agents sounds like plenty—until you have to decide which 3 fit the current repo. Some commenters say a few narrow, transparent agents are more reliable than a big bundle of configuration you haven't read yourself. The second is hook permissions. Hooks run real commands on events such as tool calls, so installing 20 hooks means letting 20 scripts run on your machine. One reply warns that before installing a large bundle, open the raw files and inspect the hooks; VirusTotal has found malicious skills hiding instructions after large blocks of blank lines and inside HTML comments, which GitHub previews may not show.

More worth reading is a tutorial from a quantitative research team. Hanako built 5 roles with Claude Code: lead is responsible for planning and deciding the next step, scout first checks data coverage, sweeper runs parameter grids, skeptic specifically attacks the results, and librarian finds strategies with sources. The team also shares a research.md that records the question, conclusions, sweep results, the weakest point found by the skeptic, and the next step. One line in the tutorial is especially blunt: generating ideas is cheap; killing bad ideas is what saves money. That's why the skeptic uses the strongest model. Some replies also zero in on this role, saying the skeptic agent is the essence.

The tutorial itself connects to QuantGPT and includes data, backtesting, and trading entry points, so it is a product tutorial. What is truly reusable is not the product name but the structure: first have one agent check the data, then have another run tests, then have the strongest one be responsible for saying no, and finally record the dead ideas. Without these steps, the more agents you have, the more duplicated investigation you get, and the thicker the bill.

This contrasts with another piece of news. AGTP's summary of Anthropic's new guidelines mentions that Anthropic recommends companies remove unnecessary subagents, and some expert agents were merged back into the main loop. A single agent plus skills often beats a complex multi-agent system; one customer ran about 2x faster. One case compressed about 15 agents into 1. In Vals AI's benchmark, an agent team cost 1.8 to 5.1 times as much as a single agent, and only one score difference was statistically significant. In another test, a multi-agent Claude workflow found 66 of 70 seeded bugs, while a single agent found at most 27.

The message in this image is not complicated: subagents are still suited to parallel work, context isolation, and tasks a single agent struggles with. But if you treat subagents as the default architecture, cost will speak first.

So a more practical starting point is: pick a task you repeat every week and give it a clear endpoint, such as a reviewed PR, a researched brief, or a failing test that goes from red to green. Start by installing two or three agents you can understand, read every hook, and keep a research.md or equivalent ledger. After a run, look at how much of the output actually remains and what it cost. Scaling up the number of agents only counts if it changes that number. Free downloads are not free—especially when four agents duplicate the same investigation.