Language Report · 002 · Developer Discourse
A reading of 11,951 Hacker News documents across four lanes of the agent conversation: the general AI-agent label, agentic systems, the Model Context Protocol, and coding agents.
Splitting the conversation's 5,906 stories into engagement quartiles puts the top quartile at a median of 9 points and the bottom at 1. The language that separates the top quartile is retrospective and first-person: the vocabulary of having built the thing and reporting back.
| Term | Top-quartile stories | Bottom-quartile | Lift |
|---|---|---|---|
| started | 131 | 35 | 3.5x |
| started building | 18 | 1 | 5.3x |
| saw | 29 | 3 | 5.3x |
| knew | 18 | 1 | 5.3x |
| decided | 51 | 12 | 3.6x |
| initially | 15 | 0 | 6.0x |
Announcement language does not travel. The conversation rewards the writer who started, saw, knew, and decided, in that tense.
The builders moved to the specific vocabulary and left the generic label to the marketplace.
Distinctive-term contrast across 11,951 documents, four lanes
The Model Context Protocol lane holds the most distinctive technical vocabulary in the conversation: its core terms run 140x to 254x more concentrated than in the surrounding discourse across 2,746 documents. Its story volume rose from 34 in the last quarter of 2024 to 589 in the first quarter of 2026. A protocol that young sitting that central is the conversation reorganizing around an interface.
The general AI-agent lane's most distinctive vocabulary is commercial: years of experience, full-stack, cloud-platform names, hiring language. The specific lanes carry the build vocabulary: MCP server, coding agent, agentic harness. The specificity shift the firm measures in hiring language runs through developer discourse too, and it points the same direction: precise vocabulary is where the practitioners are.
Same method as the firm's published wearables study: differential term analysis over a public corpus with engagement outcomes attached to every story. Fuzzy-match noise in the agentic lane (magnetic, genetic) was identified and excluded before analysis.
The engine runs continuously on any conversation a client names: a product category, a competitor, a protocol. It reports the language that travels in that conversation and the language that dies. Copy written from the first list gets engaged with.
Source: Hacker News via the official Algolia API, January 2023 through August 10, 2026; 11,951 deduplicated documents, 5,906 stories. The API returns the most recent 1,000 results per query, so the highest-volume lanes sample the newest documents; cross-lane trend comparisons are not made for that reason, and the current quarter is partial. Engagement split: top versus bottom story quartile by points and comments, 1,476 stories each.