Open Source Radar — September 5: the agent's second shift
Today's board is about everything that happens around the code — the agent that remembers you, the reference library it reads before it writes, the diagram it draws to prove its work, the patent paperwork it files, and the model it runs locally when the API bill gets annoying. Five picks off the daily board, including the two heaviest repos on it right now.
Hermes Agent (Python, ~241,800 stars)
Nous Research's self-improving agent is the biggest thing on the daily board, and its claim is the one nobody else is making: a closed learning loop. It writes new skills after hard tasks, sharpens them as you use them, and searches its own past sessions with summarization instead of making you re-explain context every morning. It lives wherever you already talk — Telegram, Discord, Slack, WhatsApp, Signal, or a terminal, all from a single gateway process — and it takes scheduled work in plain language, so daily reports or nightly backups just run. The part that separates it from another chat wrapper is the execution layer: spawn isolated subagents for parallel workstreams, and run the whole thing on local, Docker, SSH, or serverless sandboxes that hibernate when idle and wake on demand. Version 0.21.0 shipped on August 31, and it's MIT-licensed — the pick for anyone who wants an assistant that's measurably better at their work in a month than it is today.
anthropics/skills (Python, ~174,400 stars)
Anthropic's public catalog of 71 Agent Skills is the second-heaviest repo trending today, and the interesting part isn't the collection — it's what got committed this week. Two skills were updated in four days: the claude-api reference, refreshed September 1 for the current model lineup and managed agents, and frontend-design, updated September 3 specifically to steer Claude away from generic design defaults. That tells you where the category is heading — skills are being maintained like production dependencies with changelogs, not like prompt snippets people paste into a chat box. Each one is a self-contained folder of instructions, scripts, and resources that an agent loads only when the task calls for it, which is the mechanism that keeps a large capability library from eating your context window. Most of the catalog is Apache 2.0, though the document-creation skills that power Claude's file output carry their own terms — read those before you ship anything on them.
diagram-design (HTML, ~31,500 stars)
Thirty-eight editorial diagram types packaged as an agent skill, and the repo's own description tells you exactly what it's rebelling against: self-contained HTML and SVG, no shadows, no Mermaid slop. Each diagram type ships with its own rules rather than a generic template — the deep research report layout, for instance, prescribes the section order and hierarchy, the visual style, the composition, and an explicit list of anti-patterns, plus a verification checklist the agent runs before it hands you the file. That's why it's trending: most agent-made diagrams are decorative, and this one is built so the output can be checked against a spec. It works with Claude Code, Codex, and Pi, and the output is a single file you can open, present, or commit. Use it when an agent needs to explain a system in a form a human will actually review, not skim past.
patent-disclosure-skill (Python, ~7,450 stars)
The most specific thing on today's board, and the one with the best origin story: the maintainer's README opens with the complaint that after years of core R&D, his name has never once appeared in the inventor column. It's a set of six agent skills for the Chinese patent pipeline — mining patentable points from code and design docs, drafting invention, utility model, and design disclosures, rewriting a disclosure into a full application, reading published patents into plain-language notes, drafting responses to examiner objections, searching bibliographic records, and tracking shifts in examination policy. Crawling runs on Scrapy, search on Playwright, and the output is diagrams plus an editable Word file, which is what a patent attorney actually asks for. MIT-licensed and built on the open Agent Skills standard. Seven thousand stars for one country's patent bureaucracy is the tell for how narrow agent skills are about to get.
Magnitude (TypeScript, ~2,980 stars)
The youngest project here — it's three months old — and the only one trending on the daily and weekly boards at once. Magnitude is an open source inference server that profiles your machine, recommends the local models that actually fit it with estimated throughput, then downloads, tunes, and runs them, plugged into the agent you already use: Pi, OpenCode, Hermes, Claude Code, Codex, Cline, and several others. It handles speculative decoding and concurrency settings per machine, loads models on request, and unloads them when they're idle or memory fills up. Apache 2.0, macOS and Linux with Windows through WSL. The appeal is obvious once you've watched a long agent session burn through a context window: no token costs, no rate limits, and your prompts and files never leave the machine. If you've been meaning to try local models but didn't want to spend a weekend on quantization formats, this is the shortest path.
Worth watching this week: the skill layer is turning into a maintenance discipline — versioned, changelogged, and narrow enough to cover one country's patent office. The agents are becoming interchangeable; what they know is not.
Which of these would change how you work the most — an agent that learns your job, a library it keeps current, or models that run on your own hardware? Tell us in the comments.
Sources: Hermes Agent (GitHub) · anthropics/skills (GitHub) · diagram-design (GitHub) · patent-disclosure-skill (GitHub) · Magnitude (GitHub)
