I spent my morning scrolling through the Juejin GitHub trending roundups from the back half of 2025, and the thing that jumped out harder than the usual Cursor-versus-Claude-Code shouting is that agent infrastructure is quietly eating the trending page. The IDE conversation is loud because it's where engineers spend money, but the repositories that keep popping up — Agent-S from Simular AI, mem0, supermemory, TradingAgents-CN, winboat — are mostly not editors. They're pieces you bolt onto an agent to give it memory, hands, a desktop, or a domain specialty. To be fair I've been saying "agents are the next thing" for two years and I want to actually use that phrase less, but the trending signal is doing real work this time and I don't think it's noise.
The one I keep coming back to is Agent-S. The pitch is an agent that operates a desktop the way a person would — clicking through GUIs, dragging files, driving Electron apps — and the OSWorld benchmark numbers from late 2025 are the first time I've seen a computer-use system feel like more than a demo. I have not stress-tested it on my own machine yet, so I'd take my enthusiasm with a grain of salt, but the architecture feels different from the "wrap a browser in Playwright and pray" projects I keep running into. The Agent-Computer Interface framing is the right one. If you squint, this is the unsexy plumbing that eventually makes the Cursor agent tab feel quaint, and that's saying something because the Cursor agent tab is genuinely good.
mem0 and supermemory are doing adjacent work on the other end of the same problem. Every agent I've ever shipped for more than a week has eventually lost its mind because it forgot what we agreed on Tuesday, and the memory-layer approach these projects take is the first one I've seen that feels production-shaped rather than toy-shaped. I'm a little skeptical of any framework that promises persistent context without a cost, because there's always a cost and it's usually retrieval quality at the long tail. But the GitHub stars are climbing for a reason, and the integration story with Claude Code and Cursor is real now in a way it wasn't six months ago.
The meta-pattern I keep landing on is that the agent ecosystem is splitting into two camps, and I think the people picking tools should pay attention. Camp one is the polished, opinionated, end-user product — Cursor, Claude Code, Copilot, Antigravity, 通义灵码. You pay them, you open them, you get a result. Camp two is the infrastructure layer — Agent-S for hands, mem0 and supermemory for memory, the LangChain and MCP-adjacent stuff for tool routing, TradingAgents-CN as an example of a vertical-specific multi-agent stack. Camp one is what gets the Juejin headline roundups, but Camp two is what gets the GitHub trending page, and the people shipping production agents are reading both.
Honestly I think the practical advice is the same as it's been for a while: don't pick a side, build with the polished tool you actually open every day, and treat the infrastructure projects as something you evaluate when you hit a wall that the IDE-native agent can't climb. Computer-use, persistent memory, vertical agent stacks — those are the walls I expect to hit in 2026, and the trending page is telling me who's likely to be selling the ladders when I do.
I'll reassess in three months. The last time I said that I was mostly bouncing between Cursor and Claude Code, and that part hasn't changed. What's changed is that the floor underneath them is getting more interesting, and I'm spending more evenings reading agent-framework READMEs than I used to. That feels like the right direction.
Top comments (0)