🌏 中文版
Today's Highlight
No flashy framework release today, but the three repos on this list all tackle the same underlying problem from different angles: how do you make "a fleet of agents" visible and manageable to the humans running them? Orca solves it at the UI layer — desktop plus mobile. AgentTeams solves it at the protocol layer — borrowing an existing federated chat protocol, Matrix. Different paths, same destination: managing a crowd of agents is becoming the product, not just a smarter orchestration algorithm underneath it.
Trending Repos
stablyai/orca ⭐ 96,374
GitHub · TypeScript (Electron) · MIT
- What it is: A desktop AI agent orchestrator that lets you run several CLI coding agents — Claude Code, Codex, OpenCode, Grok, and others — side by side, each in its own isolated git worktree, with a desktop app and a mobile companion to monitor progress and send follow-ups.
- Why it matters: YC-backed, and it runs on your own existing Claude/Codex subscriptions rather than bringing its own model. What sets it apart from most single-agent tools is that fanning one prompt across five agents in parallel and merging the best result is treated as a first-class feature, alongside Design Mode (click a UI element to drop its HTML/CSS and a cropped screenshot straight into an agent's prompt), native GitHub/Linear integration, and SSH-based remote worktrees — the whole thing is built around managing a crowd of agents as a product in its own right.
- Tech stack: Electron desktop app + git worktree isolation + SSH remote execution + iOS/Android companion apps
- Getting started: Easy — download the desktop installer or run
brew install --cask stablyai/orca/orca; it uses your existing Claude Code/Codex subscriptions, no separate model API key needed.
agentscope-ai/AgentTeams ⭐ 5,772
GitHub · Go (control plane) + Node.js/Python (worker runtimes) · Apache-2.0
- What it is: An open-source "collaborative multi-agent OS" from Alibaba's AgentScope team, built on a Manager-Workers architecture. Every exchange — agent to agent, or agent to human — happens inside a Matrix chat room (with Element Web as the front end), instead of being buried in a framework's internal call chain.
- Why it matters: Using an existing federated chat protocol as the coordination substrate, rather than building a custom dashboard, makes agent collaboration visible, auditable, and interruptible by default. The more concrete payoff is the security model: Workers only ever hold a consumer token, while the real API keys and GitHub PATs stay inside the Higress AI Gateway — so a compromised Worker still can't touch the actual credentials.
- Tech stack: Kubernetes-native controller (Go) + Matrix (Tuwunel) IM server + MinIO shared file system + Higress AI Gateway
- Getting started: Medium — a single
curl | bashgets a local Docker setup running; production or multi-user deployments go through a Helm chart onto Kubernetes.
wheresryan22/anatomy ⭐ 15,866
GitHub · JavaScript (Node built-ins only) · MIT
- What it is: A Claude Code-compatible agent skill (also works in Antigravity, OpenCode, Codex, and MiniMax Code) that turns a technical concept into a working isometric mechanical illustration — for example, "a cache key that skips unnecessary work" becomes a pin-tumbler lock that only turns when every pin lines up.
- Why it matters: It went from repo creation to roughly 15,000 stars in about 10 days, a sign that agent skills are moving beyond text and code output into producing polished visual content — and it's deliberately built to work across multiple agent runtimes rather than locking into one platform.
- Tech stack: Pure Node.js built-in modules (zero dependencies) + SVG/WebGL + optional React/Next.js components
- Getting started: Easy —
npx skills add wheresryan22/anatomy -ginstalls it in one line, no API key required.
Notable Releases
No notable framework updates today. Both langgraph (1.2.16) and claude-code shipped a patch release within the last 48 hours, but both were bug-fix only, with no breaking changes or major new features worth covering here.
Today's Takeaway
I used to think the hardest part of multi-agent collaboration was making the orchestration algorithm itself smart enough. But Orca and AgentTeams, coming from completely different angles, both point at a more basic problem instead: how do you make a crowd of agents visible and controllable to the human running them? Orca solves it in the UI layer — desktop plus mobile. AgentTeams solves it in the protocol layer — reusing an existing chat protocol instead of inventing a new coordination algorithm. Neither one spent its effort on cleverer orchestration logic, which suggests this visibility problem may matter more right now than algorithm design does.