Today's trending list is about plumbing rather than demos: the tools that make coding agents act like a team, keep a 2.78-trillion-parameter model off the GPU, and hand an agent data it can actually trust. Four worth a look, none of them repeat coverage from this week.

gstack (TypeScript, about 139,000 stars) is Garry Tan's own Claude Code setup, published as MIT-licensed Markdown — 23 tools that reassign the agent into roles with names you'd recognise from a real org chart: a CEO that reconsiders the product, an engineering manager that locks the architecture, a designer whose job is catching AI slop, a reviewer hunting production bugs, QA that drives a real browser, and a release engineer that ships the PR. The author is the president and CEO of Y Combinator, and the repo is now the fastest-growing thing in the daily trending list by a wide margin. The interesting part isn't the prompts, it's the division of labour: instead of one enormous instruction file, each role gets a narrow brief and a handoff, which is how you stop an agent from trying to do design review and release management in the same breath. Forking it costs nothing but a clone, and it needs Bun on the machine.

kimi-k3-in-c (C99, about 10,600 stars) runs Kimi K3 — 2.78 trillion parameters, a 1.56 TB checkpoint — on one CPU, with no BLAS, no framework and no GPU. The whole engine is 176 KB of portable C, and peak memory measured 8.24 GB on an ordinary 8 GB laptop. Read the latency before you get excited: that same laptop produces one token every 26.5 seconds, because the entire model streams off disk on each step; a 128 GB workstation holding the weights in memory gets down to 5.6 seconds per token, with byte-identical output. Nobody is replacing an API key with this. What it does prove is how much of the "you need a data centre" story is really a memory-bandwidth story — MoE routing plus 4-bit weights means you touch a fraction of the model per token, and one person wrote that down in readable C.

Auto Ontology (Python, about 6,100 stars) is NVIDIA's Apache-licensed context layer for agents that query databases. The pitch is aimed at the failure everyone hits: an agent can retrieve a document and write SQL, but a revenue figure answered quickly is worthless if the definition of revenue was wrong. Auto Ontology builds a semantic graph from enterprise metadata, reconciles conflicting metric definitions across systems, checks generated queries against certified definitions and lineage, and serves the result to any agent harness over MCP, a CLI or an API, with provenance returned alongside each answer. Natural-language questions run against Postgres through NVIDIA NIM, and it ships a web UI plus Helm and Docker deploys. One caveat: NVIDIA says the repo is not accepting external contributions.

Horizon (Python, about 12,800 stars) is a self-hosted news radar that behaves like a small newsroom rather than an RSS reader. It pulls RSS, Hacker News, Reddit, Telegram, X, GitHub and financial feeds, scores every item against a rubric you define per "profile" — tech news gets impact and community discussion, engineering blogs get background and practical takeaways — merges duplicate stories, and caps each category so one busy topic can't crowd out the rest. Output is a daily briefing in English or Chinese, delivered as Markdown, a Pages site, email or a webhook, and the pipeline is also callable stage by stage over MCP so another agent can drive it. If you have ever built a folder of feeds and stopped reading it after a week, this is the missing filter.

If you've tried one of these on real work — especially gstack's role split — tell us whether the handoffs held. Write to us via our contact page.