Two of GitHub's top three trends are agent harnesses, not models

What happened

GitHub Trending this week has an unusually clean tell. The top three repos by raw star score: openclaw/openclaw at 283.1k stars ("Your own personal AI assistant. Any OS. Any Platform. The lobster way."), torvalds/linux at 221.6k, and affaan-m/everything-claude-code at 115.1k ("The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond").

Linux is the control variable — it's been in the top ten basically forever. The other two are the signal. Two of the top three trending repos this week are agent harness layers, not models, not apps, not frameworks in the traditional sense. They sit on top of whatever LLM you point them at and impose structure: skills, memory, security policies, tool routing, research conventions.

This is a different shape from the trending lists we've seen for the past 18 months. The 2024 pattern was a model release (Llama, DeepSeek, Qwen) hits, racks 60-80k stars in a week, and falls off. The 2025 pattern through Q1 was agent frameworks — LangGraph forks, CrewAI clones, ad-hoc Cursor configs. What openclaw and everything-claude-code represent is the next layer: a portable, opinionated wrapper that works across model vendors.

Why it matters

The agent ecosystem has been in a weird state. Every team building seriously on Claude, Codex, or Cursor has rolled their own version of the same pieces: a way to register reusable "skills," a memory abstraction, a permission/security boundary for tool use, a convention for when to do research vs. when to act. This is the prompt-engineering equivalent of every team writing its own logging library in 2008.

The harness layer — skills, instincts, memory, security policy, tool routing — is becoming the new package ecosystem, the way npm became the new package ecosystem for JavaScript in 2012. The everything-claude-code README pitches itself as cross-compatible across Claude Code, Codex, Opencode, and Cursor, which is the giveaway. The value isn't tied to a model vendor. It's tied to the harness contract.

openclaw's "lobster way" framing (you can roll your eyes at the mascot; the engineering decisions matter more) leans the other direction — OS- and platform-portable personal assistant, model-agnostic from the start. Different sales pitch, same architectural bet: the model is becoming a commodity input to a harness that holds the actual product value.

The community signal here matches. Both repos cleared 100k stars without a single coordinated launch event — no Hacker News front-page coup, no Product Hunt push, no celebrity tweet. That's organic developer adoption, which is the only kind of star count worth anything. By comparison, this week's earlier trending winners (freeCodeCamp at 437.9k, public-apis at 411.9k) were curation repos — markdown tables of links. Useful, but a different category. This is the first time agent infrastructure has cracked GitHub trending without leaning on a curation gimmick or a model launch.

There's a second-order point about Linux being the middle entry. The kernel sitting between two agent harness repos is the kind of accidentally perfect ranking that captures where the field is. The bottom layer is still the kernel. The top layer is increasingly the harness. The model is somewhere in between, and increasingly fungible.

What this means for your stack

If you're at the start of an agent project right now, the cost of writing your own skill registry, memory layer, and security model has just dropped sharply. Pulling in everything-claude-code or openclaw and configuring against the conventions they impose is a perfectly defensible choice — the same way `npx create-react-app` was a defensible choice in 2017, even if you knew you'd outgrow it. The default answer for "how should I structure my agent's skills directory" is no longer "figure it out yourself."

If you already have a custom harness in production, the question is whether the conventions in these repos are close enough to yours that migrating saves more than it costs. Look specifically at: skill registration format (does it match how you've defined tools?), memory abstraction (file-based, vector, hybrid?), security boundary (does it enforce least-privilege on file system and network access by default?), and the research-first development pattern (does it model the "investigate before edit" loop you've already built?). If three of four match, the migration is probably worth it. If only one matches, you'd be paying integration tax for marginal benefit.

The model-vendor lock-in conversation also shifts. A harness that genuinely works across Claude, Codex, Cursor, and Opencode means switching costs on the model side approach zero for a meaningful class of workloads. That's good for negotiating leverage, good for resilience when a vendor has an outage, and bad for any vendor's pricing power. Watch which side adapts first.

Looking ahead

The next 90 days will tell us whether this is a real layer or a temporary spike. The tell is whether the major model vendors ship official harness conventions (or buy/bless one of these projects), and whether a meaningful share of new agent code in the wild starts importing skill packages instead of hand-rolling them. If both happen, the harness becomes durable infrastructure on the order of npm, pip, or cargo. If only the second happens, expect a fragmentation period before convergence. Either way, the answer to "what does a senior dev's agent toolkit look like in 2027" is being written in these repos right now — and it doesn't look much like a model name.