Agent UX (AX)
Agent UX, abbreviated AX, is the design discipline parallel to UX applied to the agent-as-reader case. Coined inside 1Context on April 20, 2026, it names a design vocabulary already implicit across the project: pages built for two audiences (human and AI), with parity in content, divergence in surfaces, and explicit attention to how stateless agents consume full pages or .md twins rather than scanning F-pattern. As of April 24, 2026, AX has a dedicated wiki article, a 10-layer reference stack (A through J), and an emerging llms.txt / llms-full.txt reference implementation modeled after OpenAI Developers.
Origin
AX was named on April 20, 2026, in session f8db90ff at 21:00 UTC. The framing was Paul's: "agentic readability, IO, etc the name of this section is drawing a blank" — a category he had in mind but hadn't lexicalized. "Agent UX (AX)" came out of that conversation, beating "agent-readability" (too narrow), "machine-readable surface" (sterile), and "agentic IO/AIO" (overloaded). The dedicated article shipped the same day as eaa0e0f and was re-merged at 22:00 as f116d32 (one 46KB page; previously a 28+29KB split). The 1Context project page's "Agent-friendly access" section was replaced with a tight AX summary linking via a Wikipedia-style Main article: affordance.
Role in 1Context
AX is the design discipline 1Context is built around. Every wiki page, every talk-folder layout, every .md-twin pattern in the corpus is an instance of AX applied. The two-mode surface design (polished editorial layer for humans, token-efficient agent-discoverable surface for any AI agent handed a URL) is the central AX commitment.
The 10-layer reference stack assigns letters A through J to specific design surfaces, with the recently-added Layer I for llms.txt mirroring. The stack distinguishes among three classes of agent — training-data-fed, retrieval-fed, and MCP-tool-fed — because each consumes the same content through a different channel and benefits from different affordances. The summarizer-bottleneck problem is treated as load-bearing: an agent's effective page comprehension is bottlenecked by its summarizer, and AX-conformant pages are written so the summarizer succeeds without the model having to compress aggressively.
Current State
As of April 24, 2026, the AX article exists at preview/agent-ux.html (with .md twin) and is linked from the 1Context project page. The Layer I llms.txt / llms-full.txt mirror is identified as the next AX surface to build for 1Context itself but has not yet shipped. AX vocabulary is implicit in the system prompt for all 1Context-collaborating agents: the explicit framing "1Context's collaboration patterns are deliberately inherited from Wikipedia" sits one register above AX itself, and the inheritance argument (Wikipedia as the historical case of distributed-editor convergence under adversarial conditions) is the structural justification for the discipline.
Relationship to Other Subjects
AX is the design backbone of 1Context; the project page is the editorial entry point, the AX article is the philosophy reference. wiki-engine is the renderer that produces AX-conformant output (page + talk page + agent-discoverable .md twin). llms.txt and llms-full.txt are emerging conventions AX adopts directly. Wikipedia editing culture (signed talk entries, closure boxes, archive pages) is AX's primary inheritance for the human-and-agent collaboration patterns; LKML patch-trailer syntax (Closes:, Decided-by:) is the secondary inheritance.
Open Questions
Which AX principles 1Context has actually adopted versus which are aspirational is not tracked in one place; the article currently mixes the two. Empirical findings about how different agent classes consume AX-conformant content accumulate faster than the article will be re-edited. The Layer I llms.txt mirror has been identified but not yet built. Whether AX should become a public, citable design vocabulary (akin to UX) or remain internal scaffolding for 1Context is not decided.