跪拜 Guibai
← All articles
AI Programming · GitHub · Open Source

Coding Agents Get Their Own Infrastructure: 8 Trending Projects, 1.35M Stars, 14 Months

By 雨夜寻晴天 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Choosing agent infrastructure now demands a new checklist: evaluate issues, release cadence, and permissions, not star counts. The ceiling on AI coding quality depends on how teams structure context and process, making skill packs the new risk surface and the new competitive layer.

Summary

GitHub Trending's top 20 on Nov 19, 2026, was dominated by 8 projects building infrastructure for coding agents. Combined they hold about 1.35 million stars, and all were created within the last 14 months. Star counts were pulled directly from the GitHub API; the projects were filtered out if they produced business value themselves. The rest split into three abstraction layers: L1 skills and methodology (superpowers, ECC, Claude-BugHunter, modern-web-guidance), L2 orchestration (orca, t3code), and L3 harness and runtime (dsh-Go, rea). All sit on top of the same base: Claude Code, Codex CLI, and Cursor CLI.

Three preconditions made this boom possible: CLI agent interfaces converged on the same primitives, MCP unified tool protocols, and the bottleneck shifted from model intelligence to context quality. The article includes a concrete skill-pack structure and warns that stars don't equal usage, that skill packs are a supply chain attack surface, and that context bloat from too many skills kills agents. It proposes a three-step adoption plan: copy the directory structure, verify permissions, and run a controlled experiment with and without skills.

The judgment is that competition is not at the model layer but at standardizing context and process. That depends on three preconditions: top-model gaps continue shrinking, CLI interfaces stay compatible, and MCP avoids fragmentation. If any breaks, half of these projects lose their reason to exist. The previous AI wave sold shovels (compute and models); this one sells instruction manuals for shovels.

Takeaways
— Out of GitHub Trending's top 20, 8 projects are building infrastructure for coding agents, with a combined 1.35 million stars, all created within 14 months.
— The largest project, morluto/rea, has 468K stars and was created in April 2026, out pacing many established frameworks—evidence that stars currently measure hype more than maturity.
— The projects fall into three layers: L1 skills/methodology (superpowers, ECC, Claude-BugHunter, modern-web-guidance), L2 orchestration (orca, t3code), L3 harness/runtime/transport (dsh-Go, rea).
— L1's real technical content is not code but the engineering of prompt engineering, specifically the `description` field in `SKILL.md`, which acts as a routing rule for the agent.
— Three preconditions enabled this boom: CLI agent interface convergence, MCP's unification of tool protocols, and the bottleneck shifting from model intelligence to context quality.
— Star counts should not be treated as usage; evaluate projects by Issues activity, Release frequency, and last commit.
— Skill packs are a new supply chain entry point: agents execute their instructions with your permissions, so review them like scripts that run on your machine.
— Installing more than 3–5 skill packs per project causes context bloat; the agent drowns in its own instructions.
— Team adoption steps: copy the directory structure not the content, confirm permission boundaries with three questions, and run a controlled experiment comparing tasks with and without skills.
— The article's conclusion: competition is at standardizing agents' context and process, not at the model layer, dependent on three preconditions.
Conclusions

The 468K stars on rea, a repo created in April 2026, outpace most established frameworks; at this stage stars measure hype, not maturity.

Skill packs are a new supply chain vector because their instructions are executed by agents with the user's permissions; reviewing them must be as strict as reviewing a shell script.

The article's three preconditions are a falsifiable test: if top-model capability gaps widen, CLI interfaces fragment, or MCP splits into incompatible protocols, half of these projects become obsolete.

Most of the 8 projects are single-person maintenance and under a year old; teams betting on them should check for CI, issue templates, and recent merged PRs.

Concepts & terms
Skill pack
A directory (e.g., .claude/skills/) containing a SKILL.md file with metadata and instructions, plus checklists and examples. Agents load it on demand based on the description field; the description acts as a routing rule.
MCP (Model Context Protocol)
An open protocol that standardizes how AI agents connect to external tools and data sources. A single MCP integration works across multiple agent platforms, replacing per-platform native integrations.
Harness
The runtime environment that provides a coding agent with filesystem access, shell execution, and tool interfaces. It's the layer where agents actually run, as opposed to the model that powers them.
Orchestration
Coordinating multiple coding agents working on the same repository, often by isolating each agent in a separate git worktree to prevent write conflicts.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗