The fastest-growing AI repositories on GitHub this autumn are mostly not models and not frameworks. We re-read 23 of them from the GitHub API on 7 November 2026 and checked them against GitHub's own Trending windows. Four layers stand out. Agents are getting packaged behaviour, harnesses are becoming a product category, design judgement is being written down as rules, and agent tooling is moving into work that has nothing to do with writing application code.
The list, re-measured
Stars are API totals on 2026-11-07. The ten-day change compares them with our previous reading on 2026-10-28. "New" marks repositories that entered the New Runtime GitHub radar in this edition.
| Repository | What it is | Stars | 10-day change | | --- | --- | ---: | ---: | | morluto/rea | One MCP server for agent-driven reverse engineering of apps and binaries | 439K | new | | mattpocock/skills | Engineering skills from one developer's .agents directory | 317K | +12.1K | | obra/superpowers | Agentic skills framework and development method | 314K | +7.2K | | affaan-m/ECC | Harness tuning: skills, instincts, memory, security | 299K | +7.8K | | deepseek-ai/deepseek-harness | DeepSeek's open, plugin-first agent harness | 275K | new | | multica-ai/andrej-karpathy-skills | One CLAUDE.md with four coding principles | 230K | +4.8K | | anomalyco/opencode | Open-source coding agent | 225K | +2.7K | | DietrichGebert/ponytail | Skill that makes the agent write less code | 186K | +7.7K | | Graphify-Labs/graphify | Codebase to queryable knowledge graph, no vector store | 144K | +6.6K | | stablyai/orca | Fleet of coding agents on your own subscription | 118K | +9.9K | | thedotmack/claude-mem | Persistent context across agent sessions | 114K | +10.4K | | Leonxlnx/taste-skill | Art-direction skill for coding agents | 111K | new | | Panniantong/Agent-Reach | One CLI for reading social platforms and the web | 110K | +4.8K | | pbakaus/impeccable | Design vocabulary, critique and audit for agents | 105K | +14.2K | | debpalash/VoiceStudio | Local voice cloning, dubbing and transcription | 103K | +11.6K | | tt-a1i/archify | Interactive architecture diagrams from plans or code | 94K | +3.7K | | heygen-com/hyperframes | Write HTML, render video, built for agents | 80K | +6.2K | | cathrynlavery/diagram-design | 45 editorial diagram types as HTML and SVG | 66K | +3.6K | | bilawalsidhu/gods-eye-view | Public spatial feeds on a 3D globe with a voice agent | 63K | new | | cloudflare/security-audit-skill | Multi-phase security audit with verified findings | 65K | +14.3K | | HKUDS/DeepTutor | Self-hosted agentic tutor with memory | 56K | new | | pingdotgg/t3code | Control surface over the agents already on your machine | 50K | +19.6K | | Nutlope/hallmark | Design skill with tiered quality gates | 50K | +10.5K |
Behaviour ships as a package
Five of the eight largest projects are instructions for an agent rather than software: Matt Pocock's skills, Superpowers, ECC, the Karpathy-derived CLAUDE.md and Ponytail. The Karpathy repository is a single file with four principles: think before coding, simplicity first, surgical changes and goal-driven execution. It sits at 230K stars.
The reason this works is portability. Claude Code, Codex, Cursor and OpenCode all read skill folders and agent instruction files, so a set of habits written once moves between harnesses. Ponytail publishes its own benchmark: 39 tasks in Claude Code, five runs each, with 53% less code and risky logic shipped with a test 98% of the time against 68% without the skill. It is the authors' measurement and has not been reproduced here. It still shows what these packages compete on: measurable changes in agent behaviour, not features.
The harness is a product category now
DeepSeek Harness was created on 13 August and stands at 275K stars. A model lab shipping its own open harness, where every capability is a plugin, means labs now compete at the layer that runs the model as well as on the model itself. The README is explicit that this is a developer preview and that compatibility will break.
Above the harnesses, a control layer is forming. T3 Code drives the Claude Code, Codex, Cursor, OpenCode and other CLIs already installed and signed in on a machine, including from a phone. It grew from 30.9K to 50.5K stars in ten days, the steepest relative climb on the list. Orca runs a fleet of coding agents on the user's own subscription. Both treat the harness as interchangeable and the session, the subscription and the fleet as what the user actually manages.
Taste is being written down
Six of the 23 exist to stop agent output from looking generated: Impeccable, Taste Skill, Hallmark, Archify, Diagram Design and HyperFrames. They share a method. Each picks one design system or theme, uses real reference material, and then checks the result in a browser against explicit rules. Hallmark sorts its gates into a floor that never ships (contrast, focus, reduced motion, invented facts, broken phone layouts) and defaults that a build may overrule with a reason. Impeccable's detect command scans a directory or a rendered URL and can emit GitHub Actions annotations.
Impeccable gained 27.6K stars in GitHub's monthly Trending window and Hallmark gained 19.8K.
Agents leave the editor
The remaining group applies the same agent stack to other jobs. REA puts reverse engineering of applications, runtime behaviour and native binaries behind one MCP server. Cloudflare's security audit skill runs a multi-phase audit and verifies each finding independently, and it gained 38K stars in a month. God's Eye View combines live aircraft, ship, satellite, weather and public-camera feeds on a 3D globe that a realtime voice agent can query. DeepTutor is a self-hosted tutor with retrieval, memory and connected coding agents. VoiceStudio is a fully local voice-production suite.
What the stars do not say
Stars measure attention. REA has 439K stars, and GitHub's monthly Trending window credits it with 431K of them in the last month. Growth that steep attracts copies: REA's README warns that lookalike repositories exist and names morluto/rea, rea.tools and the rea-agents npm package as the only official sources. No repository on this list was installed or run for this piece. The descriptions come from each project's README.
What to try, cheapest first
- Five minutes. Add the Karpathy CLAUDE.md to a project (
curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md). In Claude Code, add Ponytail with/plugin marketplace add DietrichGebert/ponytail, then/plugin install ponytail@ponytail. - If your agent builds interfaces, pick one design skill rather than stacking several, because each brings its own rule set:
npx impeccable installornpx skills add nutlope/hallmark -g. - Large codebase: Graphify (
uv tool install graphifyy) builds the graph from the AST, deterministically and without a vector store. - Before a release: run Cloudflare's security-audit-skill.
- Running several agents at once: T3 Code or Orca.
- Watch, do not migrate yet: DeepSeek Harness, until the developer preview settles.
All 23 repositories are in the GitHub radar with their theme, growth windows and the edition that admitted them.