We Called This List Settled; Half of It Turned Over in a Week

Between September 14 and September 21, 2026, the half of this list we called settled came apart. Five of last week's ten names are gone, two of them — Ponytail and Skills — from every board we have published. The top slot changed hands for the fifth time in five issues, again to a repository that had never ranked here. What is new sits underneath: three of the fifteen projects across these two boards barely moved for a fortnight or more and then went vertical inside a week, having sat in the directory the whole time doing nothing.

#1 cloudflare/security-audit-skill — a structured security auditing framework for coding agents that uses adversarial validation and coverage tracking

+14,707 stars over 7.7 days → 17,984 · security-audit-skill star curve

Earlier: 3,024 stars on 2026-08-23

Cloudflare's skill is the clearest case of that: a flat line for more than two weeks, then a climb that has not stopped. Rather than scanning for known patterns, it opens by mapping where a codebase decides to trust its inputs, hunts against that map so it can report what it never examined, then turns a second fleet of agents loose on its own findings with instructions to disprove them.

#2 alibaba/open-code-review — an AI-powered code review CLI that combines deterministic engineering with agents for high-precision feedback

+13,257 stars over 6.3 days → 38,368 · open-code-review star curve

Earlier: 21,209 stars on 2026-08-23

Alibaba's entry aims at the two failures that make agent review unusable in practice, not at the quality of the review. One is position drift, where a comment is right about the code and wrong about which line it hangs on. The other is silent truncation, where the agent reviews as much of a changeset as fits and says nothing about the rest. Deterministic plumbing handles both, leaving the agent the judgement.

#3 hypit-ai/hypit — Hypit – Open‑source AI agents that clone and remix viral videos at scale

+11,547 stars this week → 11,872 · hypit star curve

Earlier: 325 stars on 2026-09-14

Hypit is the steepest climb on the main board by a wide margin, and was effectively unknown seven days ago. You hand it a reference video or a description; it writes an entire production — footage, captions, B-roll, effects — as a declarative document and renders it in headless Chromium, and it can do all of that without calling an external model at all.

#4 deepseek-ai/deepseek-harness — a modular agent harness with an everything-is-a-plugin architecture

+8,642 stars this week → 231,191 · deepseek-harness star curve

2026-W35: +42,534 · 2026-W36: +16,749 · 2026-W37: +8,162 · 2026-W38: +8,562 · up and down across 5 issues · Earlier: 185,022 stars on 2026-08-23

DeepSeek Harness is one of two names left that have appeared in every issue we have published. Two weeks ago we called the end of its long slide noise rather than a floor. It has now landed slightly ahead of itself twice running — still no recovery, but no longer a fall that happened to pause.

#5 browser-use/jev-ultrafast — Jev Ultrafast – Fast, LLM‑driven browser automation with a dynamic indexed action space

+8,617 stars over 2.7 days → 11,789 · jev-ultrafast star curve

Browser-use's newest project has been on our board under three days, enough to judge its speed and not its shape. The design is a deliberate narrowing: the page becomes a numbered table of controls, and the model is asked for one row and one verb per turn. That fits in a single round-trip to a small model, which is why a flight search finishes in seconds.

#6 tt-a1i/archify — a deterministic system mapping tool that turns agent-generated JSON into validated interactive architecture diagrams

+7,531 stars this week → 68,247 · archify star curve

2026-W36: +19,312 · 2026-W37: +12,853 · 2026-W38: +9,964 · gaining less each issue, 4 issues running · Earlier: 15,058 stars on 2026-08-23

Archify has been drifting down the board since it led in August, and it is still the one long run here that has never once reversed. The tool is a narrow bet: an agent may describe a system but may not draw it. The description goes to a validator that enforces layout and schema before anything renders, so a diagram that looks authoritative cannot be invented.

#7 JustVugg/colibri — a high-performance inference engine for massive MoE models using AI memory multitiering

+6,824 stars this week → 36,595 · colibri star curve

Earlier: 25,991 stars on 2026-08-24

Colibrì is the only project in the top ten with nothing to say about agents. Written in C, it stops treating graphics memory as the box a model must fit inside and treats VRAM, system memory and the SSD as one hierarchy, pulling whichever experts a mixture-of-experts model routes to off disk as it goes. A frontier-scale model then runs on a desktop, slowly, instead of not at all.

#8 debpalash/VoiceStudio — a local-first audio production suite for voice cloning, video dubbing, and long-form audio generation

+6,788 stars this week → 33,527 · VoiceStudio star curve

2026-W37: +7,528 · 2026-W38: +6,912 · holding the same pace, 3 issues running · Earlier: 11,235 stars on 2026-08-23

VoiceStudio is the control group. Half the names around it were replaced this week; it did not notice, which makes the churn above it look like something happening to the board rather than to the field. It is also the only entry up here built for people who make audio rather than software.

#9 affaan-m/ECC — an agent harness operating system that provides AI agents with a coordinated engineering system and toolbox

+5,976 stars this week → 263,717 · ECC star curve

2026-W37: +8,532 · 2026-W38: +6,460 · up and down across 3 issues · Earlier: 242,167 stars on 2026-08-23

ECC is now the largest project on this board by total stars, a position it inherited rather than earned — Skills, which sat above it last week, is gone. What it sells is the opposite of the narrow tools above it: not one rule to bolt onto an agent but a whole organisation of them, with a fixed cycle of planning, testing and verification to work through.

#10 stablyai/orca — an AI orchestrator for running multiple CLI coding agents in parallel isolated worktrees

+5,789 stars this week → 73,578 · orca star curve

2026-W35: +5,420 · 2026-W36: +5,522 · 2026-W37: +5,082 · 2026-W38: +5,022 · up and down across 5 issues · Earlier: 51,254 stars on 2026-08-23

Orca is the other name present in every issue. Last time we described it as replenishing at a fixed rate, neither discovered nor tired of. This week it broke that, arriving ahead of any week we have recorded for it. Nothing about running coding agents in parallel worktrees has changed — but half the board changed around it, and the two newcomers at the top are both things you would want more than one of.

Breaking out

Ranked by growth relative to their own size, not by stars added. These are small enough that a week can multiply them — and small enough to have escaped you.

tamaratran/fast-jev-compaction · +520.9% over 2.8 days → 5,147 stars · fast-jev-compaction star curve

Fast-jev-compaction is the other half of the Jev arrival. It replaces Claude Code's summarising compaction with a decision taken one tool call at a time: the transcript goes to the model with tool results stubbed out, and the model answers only which of those interactions can be dropped, leaving everything a person or the assistant wrote word for word.

latent-spaces/brag · +420.4% this week → 6,172 stars · brag star curve

Earlier: 1,108 stars on 2026-08-23

Brag is built on Hyperframes, which was in last week's top ten — a rung of the stack becoming somebody else's dependency inside a fortnight. Point it at a repository and it writes the composition brief itself, then renders a short launch video with music and motion. It belongs to the flat-then-vertical group too.

dexmal/opendm · +270.9% over 6.4 days → 2,055 stars · opendm star curve

Earlier: 314 stars on 2026-08-23

OpenDM is the only thing on either board that touches hardware: a vision-language-action framework with an open model family behind it, trained across several robot bodies rather than one, so the same policy can be fine-tuned onto a different arm. It is the third of the three that sat still for weeks and then abruptly did not.

Human-Agent-Society/reef · +207.6% this week → 3,780 stars · reef star curve

Earlier: 327 stars on 2026-09-04

Reef is the most ambitious idea in this issue and the least proven. It proposes to close the loop everything else here leaves open: inference produces feedback, feedback produces a versioned update, and the update goes to the weights and to the harness around them. Every other project here treats the harness as something a person edits.

Tencent/BrowserSkill · +206.7% this week → 6,038 stars · BrowserSkill star curve

Earlier: 1,280 stars on 2026-08-23

Tencent's bridge is the opposite bet from Jev Ultrafast. Rather than driving a clean browser through an indexed action space, it borrows the one you are already signed into, opening a separate Agent Window so your own tabs are not hijacked. One makes the browser legible to a small model; the other makes your session usable by whichever agent you already run.

The board we published last week had a settled membership and a rotating leader. Only one of those halves survived. What replaced the other did not climb into view: three of the arrivals had been going nowhere in the directory for a fortnight, and two more were not in it at all seven days ago.