Its 255,068-skill comparison shows multilingual growth, while lexical discovery remains a likely constraint on cross-language reuse.
By RuntimeWire Staff · Published
Primary source: Plicara Labs
Why it matters #
Agent skills are becoming multilingual faster than GitHub documentation did, creating a concrete discovery problem for registries, marketplaces and agent workspaces built around text matching.
Plicara Labs published an analysis of agent-skill languages on August 25th, 2026, showing that the non-English share of newly written skills rose from 13.0% in the first quarter of 2026 to 16.3% in the second. The increase gives the researchers behind Plicara an early view of an agent-development layer spreading internationally within months of its creation.
Plicara's quarter-to-quarter comparison covers 255,068 newly written skills. The non-English share rose from 13.0% in Q1 2026 to 16.3% in Q2, with 95% confidence intervals of 12.8% to 13.1% and 16.1% to 16.4%.
The researchers describe Plicara as a pre-alpha applied research lab focused on training and benchmarking AI models. Its public materials do not identify its founders, funding, headquarters, customers or revenue, and the lab says it has not released a model. Its nearer-term work centers on evaluation tools and reproducible research: Plicara has released regexbench for testing model-generated regular expressions and labloop for agent-driven experiments, both on PyPI under Apache-2.0 licenses.
That approach matters here because the headline number changes depending on where researchers look.
A young format is spreading quickly
An agent skill is a folder centered on a SKILL.md
file that gives an AI agent procedural instructions in prose. Scripts, reference documents and other resources can travel with it. The agent reads the description and decides when the skill is relevant to a task.
Anthropic introduced Agent Skills on October 16th, 2025, and later published the format as an open standard. By July 2026, the GitSkills dataset had collected 3,797,117 skill-file occurrences from 282,200 public GitHub repositories. Identical files were grouped into 1,877,981 distinct contents.
Plicara's research page reports analyzing 1,870,299 distinct skill contents. The GitSkills paper and Plicara's public analysis repository report 1,877,981 distinct contents, leaving an unresolved difference of 7,682. Plicara separately says 6,810 files in its analysis fell below the language classifier's confidence threshold.
Plicara found that English accounted for 85.3% of classified skills. Chinese was the largest non-English language at 6.2%, followed by Japanese at 1.7%, German at 1.6%, Korean at 1.2%, Portuguese at 1.1%, Spanish at 0.9% and French at 0.4%. The Chinese group contained 104,985 simplified-script skills and 9,112 traditional-script skills.
The overall 14.3% non-English share sits close to the 13.0% reported for GitHub documentation in a separate 2026 open-source language study. The composition differs: Chinese represented 6.2% of Plicara's skills corpus, compared with 3.3% of repositories in that broader documentation study.
The quarterly movement came from different language groups moving in different directions. Chinese rose from 4.2% of newly created skills in Q1 to 5.3% in Q2. A group combining German, French, Spanish, Portuguese, Italian, Russian and Dutch increased from 2.7% to 5.5%. Japanese fell from 3.2% to 2.5%, while Korean edged down from 2.2% to 2.0%.
Plicara tested whether copying or bulk up could be manufacturing the increase. Among skills copied only once, the non-English share rose from 14.7% to 18.1%. When each repository was counted once, it still increased from 14.5% to 16.8%.
Writing is globalizing while discovery may lag
A growing multilingual corpus creates a practical discovery problem for registries and agent workspaces. Plicara argues that lexical discovery is a likely constraint because searches in English are less likely to surface skills written in Chinese or other languages. That is a hypothesis about the mechanism, rather than a causal result established by the repository crawl.
One possible explanation is that discovery systems still rely heavily on language matching. Multilingual semantic search, translation and cross-language recommendations could help developers find procedures whose text does not share the language of their query.
Commercial services are already assembling a discovery layer around skills. SkillsMP emphasizes aggregation, source links, repository context and localization. Agensi markets a paid skills marketplace with creator revenue sharing and automated security scanning. Vercel said the skills.sh API queried more than 600,000 skills as of June 5th, 2026. Those services are adjacent to Plicara's research rather than established direct competitors, but they show where multilingual discovery could become a product feature.
What the dataset can and cannot prove
Language is an incomplete proxy for geography. Developers in India, Nigeria, Singapore and other countries may write their skills in English, disappearing into the largest category. Plicara's 14.3% figure should be read as a lower bound on development outside English-dominant communities, rather than a measurement of work outside the United States.
The time series also rests on a subset. Commit history was available for 24% of skills. GitSkills is a July 2026 point-in-time snapshot; its dataset documentation says files deleted between discovery and content retrieval were excluded, while files deleted after the crawl could remain in the snapshot. Plicara says the dated sample leans toward frequently copied files, although its robustness checks preserved the quarterly increase after controlling for copying and repository concentration.
Classifier choice appears to explain little of the disagreement among published language estimates. Plicara ran its identifier and the classifier used by another study over the same documents; they agreed on 97.6% of them, with their English estimates separated by 1.2 percentage points. Published English shares still range from 65.0% in a small healthcare sample to 99.7% in a crawl seeded with English queries. Corpus selection is doing much of the work.
Counts also say little about quality or safety. Snyk reported security flaws in 36.8% of a 3,984-skill sample. That study used a different corpus, but its result is a useful warning against treating a growing file count as evidence that the underlying instructions are reliable.
Plicara also found an AI co-author trailer in the commit history of 30.4% of skills. That is a floor: trailers can be removed, lost during commit squashing or omitted by tools, as Git's commit documentation makes clear. Agents are increasingly helping developers write instructions that other agents will later load.
Execution quality remains the unanswered technical question. A repository crawl can identify a language, copy history and revision pattern. It cannot establish whether a model follows an instruction equally well in Chinese, Portuguese, German or English. Answering that will require multilingual execution benchmarks rather than another file count.
Plicara's research establishes that multilingual supply is growing quickly. Its explanation for how language affects discovery remains a testable hypothesis for the registries and agent workspaces now trying to organize that supply.