AI Agent GitHub Digest — 2026-10-17

🌏 中文版

Today's Highlight

Today's trending list circles almost entirely around Claude Code as a harness itself: Anthropic's official plugin directory gained 1,552 stars in a single day, while superpowers turns it into a development methodology, ruflo wraps it into a multi-agent swarm, and ai-job-search uses it as a full job-hunting pipeline. Claude Code is shifting from "a CLI" into the foundation other people build their houses on.

Trending Repos

anthropics/claude-plugins-official ⭐ 40,252 (+1,552)

GitHub　·　Python　·　Apache-2.0

• What it is: Anthropic's own, officially maintained directory of Claude Code plugins. Internal plugins (/plugins) and third-party ones (/external_plugins) are kept in separate folders, and installing one is as simple as /plugin install {name}@claude-plugins-official.
• Why it matters: There's no new technology here — what matters is the governance it establishes. A published plugin name is an immutable slug; renaming it requires adding an entry to the top-level renames map in marketplace.json so existing installs migrate automatically. In other words, this repo turns Claude Code's plugin system from "everyone rolls their own" into a versioned supply chain. +1,552 stars today is also unusually high for a repo Anthropic runs itself.
• Tech stack: Pure config files (plugin.json + marketplace.json) plus a review process — no runtime code involved
• Getting started: Easy — browse via /plugin > Discover or run the install command directly; the bottleneck is passing review, not the tech

obra/superpowers ⭐ 300,542 (+973)

GitHub　·　Shell　·　MIT

• What it is: A "software development methodology" packaged as a plugin that installs into a dozen-plus harnesses — Claude Code, Codex, Cursor, and more. It forces the agent to first nail down what you're actually trying to build, break the spec into digestible chunks, and lay out a YAGNI/DRY implementation plan before letting subagents actually write code.
• Why it matters: It and the official plugin directory above validate each other — superpowers itself is installed as a plugin from claude-plugins-official, which shows that "skill-driven development workflow" has moved from personal habit to a packaged, cross-harness-installable product, rather than something tied to Claude Code alone.
• Tech stack: Shell scripts plus multiple SKILL.md files; installation path varies by host harness (Claude Code goes through the official marketplace, others use their own plugin systems)
• Getting started: Easy — no extra configuration after install; skills trigger automatically based on conversation content

ruvnet/ruflo ⭐ 75,533 (+672)

GitHub　·　TypeScript　·　MIT

• What it is: A "meta-harness" wrapped around Claude Code/Codex. The author's framing is "Agent = Model + Harness," and ruflo is that harness layer — managing 100+ specialized agents, coordinating swarms, running local vector memory, and even letting agents on different machines talk to each other through federation.
• Why it matters: Projects that "wrap another harness around the harness" have been showing up on trending a lot lately, a sign that a single agent in a single conversation is no longer enough — developers are starting to treat multi-agent coordination as its own infrastructure problem, instead of each project rolling its own solution.
• Tech stack: TypeScript + an npm package (npx ruflo) + an MCP server, also distributed as a Claude Code plugin
• Getting started: Medium — npx ruflo init gets a minimal setup running in one line, but swarms, federation, and the 314 MCP tools need extra configuration and reading the docs

MadsLorentzen/ai-job-search ⭐ 48,176 (+879)

GitHub　·　Python　·　MIT

• What it is: A job-hunting framework that turns Claude Code into a full-stack workflow — /scrape finds openings, /apply scores fit and drafts a tailored CV and cover letter (typeset in LaTeX), and /interview helps you prepare. The core pipeline is language- and country-agnostic; only the job-portal search skills are built for the Danish market.
• Why it matters: The author, a geophysicist laid off from his job, used this exact workflow himself — 69 tailored applications, 20 first-round interviews, and one signed contract later, he started as an AI engineer in June. It's a rare case of applying agent tooling to something as high-stakes and privacy-sensitive as a job search, with the real numbers published openly instead of just a demo.
• Tech stack: Claude Code CLI + Python 3.10+ + Bun (job-search CLI) + LaTeX (lualatex/xelatex for typesetting the CV and cover letter)
• Getting started: Medium — requires a Claude Code subscription, Bun, and a full LaTeX distribution; the job-search skills default to Danish job sites, so switching regions means editing the skill yourself

Notable Releases

Agno v3.1.4

Release Notes

• Key changes: ClaudeAgent/CodexAgent gain a sandbox= parameter that runs the entire harness process inside an isolated, per-session computer separate from the host. Session state is kept through a lease row plus checkpoints in Postgres, so a turn can keep going even after losing the container, the host, or the replica that ran the last turn. Two sandbox providers ship: DockerSandbox (UID 1000, read-only filesystem, every privilege dropped) and AgentCoreSandbox (running on AWS Bedrock AgentCore's Firecracker microVMs), plus a "warm sandbox" that keeps a session's computer around for 300 seconds by default so back-to-back turns skip the cold start (4.6s vs. 8.2s on Docker).
• Breaking changes: The sandbox provider packages moved to agno.sandbox.docker and agno.sandbox.agentcore, so DockerCheckpointTransfer imports need updating. The new require_native_resume/strict_persistence options default to stricter behavior: what used to silently fall back to a fresh conversation when session state couldn't be restored now raises a native_state_missing error instead.
• What this means for you: If your project imports classes from agno.sandbox, update the import paths. If your session-recovery logic relied on the old silent fallback to a fresh conversation, confirm the new strict checks won't block a normal flow before upgrading.

Today's Takeaway

I used to think a "sandbox" was just an isolated environment to stop an agent from breaking things. Looking at Agno's warm-sandbox and checkpoint-recovery design today made it clear the real problem being solved is harder: making a session resumable on any machine. That's the same thread running through the official plugin directory locking down plugin names and ruflo's cross-machine federation — agent infrastructure is converging on "resumable and portable," rather than each project solving its own narrow slice of the problem.

References

• anthropics/claude-plugins-official
• obra/superpowers
• ruvnet/ruflo
• MadsLorentzen/ai-job-search
• Agno v3.1.4 Release Notes
• GitHub Trending (2026-10-17)