Plicara reports that agent skills — prose instruction files used by AI agents — are becoming less English-only. In its analysis, the share of newly written skills in a language other than English rose from 13.0% in the first quarter of 2026 to 16.3% one quarter later, across 255,068 skills. The finding is single-source in this cluster and should be read as Plicara’s measurement, not an independently confirmed benchmark. But the dataset is large enough to make the reported shift worth watching: Plicara says the confidence intervals for the two quarters “nowhere near” overlap, and it contrasts the pace with GitHub-wide non-English documentation, which it says took a decade to move from 3.7% to 13.0%. The object being measured is specific. Plicara describes a skill as a SKILL.md file in a folder, written in plain prose, that an AI agent loads when it decides the instructions are relevant to a task. Plicara says Anthropic published the specification in October 2025, and that nine months later the GitSkills dataset counted 3.8 million such skills across 282,200 public repositories. That matters because skills are not software in the usual programming-language sense. They are instructions written for a multilingual model. Plicara’s argument is that there is no inherent technical reason for a developer in Shenzhen or São Paulo to write that procedure in English if the agent can read the developer’s own language. The article does not claim to measure whether agents follow every language equally well; it says that question is harder than a file crawl can answer. Plicara says it ran language identification over the prose body of every distinct skill after removing front matter and fenced code. Across the dataset, it found 14.3% of skills were not in English. For Chinese-language skills, it reports 104,985 in simplified Chinese and 9,112 in traditional Chinese. It also says 6,810 skills fell below its confidence threshold and were counted as neither English nor non-English in the detailed rows. The comparison Plicara emphasizes is not that agent skills are dramatically more multilingual than GitHub documentation overall. It cites a 2026 International Conference on Software Engineering study putting repository documentation at 13.0% non-English, close to Plicara’s 14.3% figure for skills. The sharper distinction, according to Plicara, is Chinese: 6.2% of skills in its dataset versus 3.3% of repositories in the cited documentation study. Plicara also cautions that language is not a clean proxy for geography. It notes that large developer populations such as India, Nigeria and Singapore often write in English, so a non-English count would miss much of the ecosystem outside the United States. In that framing, the non-English share is a lower bound on international participation, not a full map of where agent-skill development is happening. Plicara says other published figures do not agree with each other, and attributes the differences to different sampled populations, including curated marketplaces, domain-specific slices and crawls seeded by English-language queries. The safe conclusion is narrower: Plicara’s crawl suggests agent-skill repositories are already materially multilingual, with Chinese standing out against broader GitHub documentation baselines. Who benefits: Developers and teams that already work outside English may benefit if agent systems can consume task instructions in their strongest working language. Tool vendors that support multilingual discovery and evaluation of skills could also have an advantage. Who's exposed: Platforms or marketplaces that curate, rank or test skills primarily in English may underrepresent part of the activity Plicara says is emerging. The article does not provide performance data, so exposure around execution quality by language remains unresolved.