Agentic frameworks ate GitHub trending: openclaw at 283k, superpowers at 113k

What happened

GitHub trending this week tells a story that two years of LangChain conference talks didn't: the agentic-framework era has arrived, and it's measured in six-figure star counts. openclaw — a personal-AI-assistant framework pitched as 'any OS, any platform, the lobster way 🦞' — is sitting at 283.1k stars, behind only freeCodeCamp (437.9k) among trending repos. Right behind it, Jesse Vincent's `obra/superpowers` — billed as 'an agentic skills framework & software development methodology that works' — crossed 113.5k.

These aren't curriculum repos. They aren't awesome-lists. They're runnable code: openclaw is a local-first agent runtime targeting Windows, macOS, Linux, and (per its own claims) anywhere you can install a Node binary. superpowers is a skills layer that plugs into Claude Code and similar agentic harnesses, codifying patterns for multi-step coding work — search, plan, edit, verify — that practitioners have been re-implementing by hand for eighteen months.

The pattern matters because of what's *not* on the list. No new VS Code plugin cracked the trending top tier this cycle. No "AI for your codebase" chat wrapper. The two biggest non-curriculum repos are both about giving agents real, structured execution surfaces — runtimes and skill libraries — rather than another autocomplete UI.

Why it matters

The AI-coding stack has been bifurcating for a year, and these two repos are the clearest snapshot of where the split landed. On one side: hosted, opaque agents (Copilot Workspace, Cursor's background agents, Devin) — closed runtimes you rent. On the other: local, inspectable agent frameworks you run yourself, with skills you can read, fork, and version. openclaw and superpowers are both in the second camp, and the star velocity says the second camp is winning on developer attention if not yet on revenue.

Look at what each is actually solving. openclaw answers the question "how do I run a real agent on my laptop, on my OS, with my files, without a cloud round-trip for every tool call?" It's the descendant of every "local LLM + tools" experiment from the past two years, but with enough polish that 283k people pressed the star button. superpowers answers a different question: "once I have an agent loop, what *skills* do I plug into it?" — and codifies a methodology (the README's word: methodology, not framework) for how a senior engineer would actually break down a multi-step task.

The community signal in the comments under both repos splits cleanly. The skeptics — and there are loud ones — point out that star counts overstate adoption: stars are bookmarks, not deployments, and an agentic framework with no shipped product behind it is a graveyard waiting to happen. The believers point at the opposite data: when curriculum repos and joke awesome-lists historically dominated trending, executable agent frameworks couldn't crack the top 20. Now two of them are top-three. That's a regime change in what developers are *trying to build*, regardless of what they ship.

There's also the Jesse Vincent factor on superpowers. Vincent (obra) has been in this space long enough — Best Practical, Perl mail tooling, then Claude Code adjacent work — that the framework reflects real production experience, not a weekend hackathon. The README is unusually blunt about what doesn't work: skills that try to be too general, agents without verification loops, frameworks that prescribe instead of describing. That's the kind of opinionated documentation that makes a framework load-bearing, not a curiosity.

What this means for your stack

If you've been waiting for the agentic-framework space to consolidate before betting on one, this week is a signal — though not yet a verdict. Two things to do this quarter.

First, audit your existing "AI-in-the-loop" code paths. If you've built bespoke prompt-chain glue inside an internal tool — the "we wrote our own little agent" pattern that's now everywhere — compare your loop to what superpowers prescribes. The methodology section is short enough to read in 20 minutes, and if your homegrown version is missing the verification step, you have a known-bad pattern in production. Don't rewrite on top of someone else's framework yet; do steal the methodology.

Second, if you're evaluating local-first agent runtimes for a desktop or internal-tools use case, openclaw is now serious enough to bench. Run it against whatever you're using (LangGraph, AutoGen, your own loop) on the same three tasks: a multi-file refactor, a bug repro from a stack trace, and a "read this codebase and explain X" cold start. The 283k stars don't tell you it's better than your current stack — they tell you 283k people thought it was worth checking, and the cost of a one-day bench is small relative to picking the wrong runtime for the next two years.

The harder call: should you start *publishing* skills back upstream? superpowers is structured as a skills marketplace in spirit, even if the on-ramp is rough. If your team has an internal skill — say, "safely run a migration against staging, verify row counts, post to Slack" — it's worth asking whether that's better as a private prompt or a public, versioned skill that earns you reputation and pulls fixes from the community.

Looking ahead

The trending list will rotate, as it always does. But the bigger question — whether agentic frameworks become the new "web framework" tier of the developer stack, with consolidation around two or three winners — is the one to watch through the next two quarters. If openclaw and superpowers are still in the trending top ten in October, the answer is yes, and the next round of dev-tools VC theses will be written against that fact. If they fade, the lesson is that the hosted-agent vendors won the war for production, even as they lost the war for stars.