Skillselion Research · Edition of 2026-08-12
The Agent Skill Maintenance Census
We expected a graveyard. We read the last-push date of every repository behind 79,848 listings and found the opposite: the median install runs on code pushed 6 days ago, and abandoned code carries almost no installs at all.
On 2026-08-12 we dumped the whole Skillselion production catalog and measured the age of every listing’s last repository push. The story software registries usually tell, a thin active core on a mountain of abandonware, has not arrived here. Half of all installs sit on week-old code, the 180-day-stale tail holds 1.0% of installs, and the census found exactly 3 listings a full year without a push. You can reproduce every number from the dataset; the method and its limits are in the methodology.
Finding 01
The median install runs on code pushed six days ago
Weight every listing by its installs and take the median: the code behind it was pushed 6 days before the census. 79,875,458 installs, 48.5% of the entire ecosystem, sit on repositories pushed within the last week; 66.8% within the last month. The unweighted median listing is 26 days old, so the code people actually install is fresher than the catalog at large. Adoption and maintenance travel together.
- Code pushed within 7 days48.5%
- Code pushed within 30 days66.8%
- Covered by a last-push date at all90.6%
Finding 02
Half the catalog idles. It holds a quarter of the installs.
The tail looks different from the core. 47.9% of dated listings went >30 days without a push, but they hold 23.8% of installs. Push the cutoff to 90 days and the listings share stays large (28.2%) while the install share collapses to 2.6%. At 180 days it is 9.5% of listings and 1.0% of installs. Idle code exists in bulk; installed idle code barely exists. The largest stale listings in the census are humanizer-zh (43,355 installs, 204 days) and design-doc-mermaid (32,625 installs, 225 days), small next to anything in the live core.
- No push in 30 days: 47.9% of listings23.8% of installs
- No push in 90 days: 28.2% of listings2.6% of installs
- No push in 180 days: 9.5% of listings1.0% of installs
Finding 03
Too young for a graveyard
Exactly 3 listings out of 75,800 with usable dates went a full year without a push, and together they hold 0 installs. Software registries accumulate abandonware over decades; this ecosystem’s distribution format took off in late 2025 and has not existed long enough to decay. That cuts the other way too: nothing here has yet survived long enough to prove durability. The 180-day bucket, 9.5% of dated listings today, is the number to watch across editions.
Finding 04
The risky code is fresh, not stale
Cross this census with the security census and a comfortable assumption fails. The overlap of stale (>180 days), shell-capable and never-audited listings is 208 listings holding 7,335 installs. The security census’s full audit blind spot holds 15,858,667 installs. So over 99% of the unaudited shell-capable install mass sits on actively maintained code. Rot is not the threat model here. Velocity is.
Method
Methodology
The census is a complete dump of the Skillselion production catalog taken on 2026-08-12: 79,848 listings and 164,841,042 cumulative installs, aggregating public registry data, chiefly the skills.sh registry, alongside GitHub and other public sources. Staleness is the gap between the census date and the GitHub pushed_at of the listing’s repository. 75,800 listings (94.9%, holding 90.6% of installs) carry a usable date and form the denominator for every staleness figure; 259 rows report an epoch-zero timestamp (a registry artifact) and are excluded along with undated rows.
One structural limit to keep in mind: pushed_at is repository-level. A monorepo hosting fifty skills refreshes all fifty with one push, so per-skill freshness is overstated wherever big multi-skill repositories dominate, while staleness is reliable (an untouched repository is untouched for every skill in it). Install counts measure registry install events, not active users.
A push timestamp measures activity, never quality or safety. Skillselion is an independent project and is not affiliated with Anthropic, OpenAI, Cursor, Claude, Claude Code, or Codex.
Data
Download the dataset
The full census runs 79,848 rows, one per listing, with repository, last-push timestamp, installs, audit flag and shell-permission flag, so both the staleness buckets and the security overlap recompute from one file. It is free to use with attribution under CC BY 4.0: cite “Skillselion Agent Skill Maintenance Census, August 2026” and link this page.
FAQ
Questions journalists ask
Is AI agent code maintained?
The popular core is, aggressively: the median install in this census sits on code whose repository was pushed 6 days earlier, and 48.5% of all installs sit on code pushed within a week of the census date. The long tail moves slower (the median listing is 26 days old), but abandonment barely exists yet: listings untouched for 180+ days hold 1.0% of installs, and only 3 listings went a full year without a push.
Does a recent push mean a skill is safe or good?
No. A push timestamp measures activity, nothing else. Freshness also runs on the repository, not the file: one push to a monorepo refreshes every skill it hosts. Pair this census with the Agent Skill Security Census before treating fresh as trustworthy; most of the audit blind spot there sits on actively maintained code.
Why does the ecosystem have almost no year-old code?
Because the ecosystem is younger than that. Agent skills as a distribution format took off in late 2025, so most repositories in the census simply have not existed long enough to rot. The interesting question is what the 180-day bucket (9.5% of dated listings today) looks like by the next edition, and frozen editions exist so that comparison stays honest.
Can I reuse these numbers or the dataset?
Yes. Cite "Skillselion Agent Skill Maintenance Census, August 2026" and link this page. The raw dataset (79,848 rows with last-push timestamp, installs, audit and shell-permission flags per listing) is downloadable on this page under CC BY 4.0.