Author here. I got tired of "best AI agents" lists that rank the same
five projects from 2023, so I built something that only looks at what
shipped recently.
Every Monday a pipeline sweeps GitHub for agentic repos created in the
last 30 days — 1,757 this week — and scores them on a Momentum Score
rather than raw popularity: 35% star velocity (stars/day since launch),
20% traction (forks/watchers), 20% development activity (commit cadence,
push recency), 15% craft (README depth, license, metadata), 10% relevance
(how clearly it reads as an agent, so general AI tooling doesn't crowd
out actual agents).
Why a separate metric: I also keep an all-time Top 100 scored on
reliability, tool selection quality and community health. That formula is
useless on new projects — a large share of it is CI presence, test
coverage and contributor signals, all near-zero on a four-day-old repo.
Everything compresses into a narrow band and the ranking quietly
degenerates into star count.
Honest limitations:
- Scoring is static analysis of repo metadata, not sandboxed execution.
I can't tell you these agents complete tasks well, only that they show
the signals of a serious project with real traction. Sandbox eval is
the roadmap.
- Star velocity is gameable. I cap implausible velocity (>5k stars/day)
because renamed or newly-public repos carry an old star count against
a fresh created_at — that guard fired this week on a repo showing
11,346 stars/day.
- The 30-day window is a tradeoff: 7 days gave fewer candidates than
board slots, so the ranking meant nothing.
Everything's open: full data as JSON (CC-BY), per-agent rank history
since March, and the methodology with all weights published.
Genuine question for this crowd: what would you want measured that I'm
not measuring? Static analysis only gets so far and I'd rather hear it
from people who actually run these things.
Update 1,757 from Monday's digest on the day.
Are you sure that Fable 5.1 outperforms Astra? Feel like early indications and testing appear to show for the first time a true Agentic experience? This could open the door to a true loop build in every corner of the world.
Every Monday a pipeline sweeps GitHub for agentic repos created in the last 30 days — 1,757 this week — and scores them on a Momentum Score rather than raw popularity: 35% star velocity (stars/day since launch), 20% traction (forks/watchers), 20% development activity (commit cadence, push recency), 15% craft (README depth, license, metadata), 10% relevance (how clearly it reads as an agent, so general AI tooling doesn't crowd out actual agents).
Why a separate metric: I also keep an all-time Top 100 scored on reliability, tool selection quality and community health. That formula is useless on new projects — a large share of it is CI presence, test coverage and contributor signals, all near-zero on a four-day-old repo. Everything compresses into a narrow band and the ranking quietly degenerates into star count.
Honest limitations: - Scoring is static analysis of repo metadata, not sandboxed execution. I can't tell you these agents complete tasks well, only that they show the signals of a serious project with real traction. Sandbox eval is the roadmap. - Star velocity is gameable. I cap implausible velocity (>5k stars/day) because renamed or newly-public repos carry an old star count against a fresh created_at — that guard fired this week on a repo showing 11,346 stars/day. - The 30-day window is a tradeoff: 7 days gave fewer candidates than board slots, so the ranking meant nothing.
Everything's open: full data as JSON (CC-BY), per-agent rank history since March, and the methodology with all weights published.
Genuine question for this crowd: what would you want measured that I'm not measuring? Static analysis only gets so far and I'd rather hear it from people who actually run these things. Update 1,757 from Monday's digest on the day.
reply