tyson chan

spotted in the wild

OpenAI shipped model-composed interfaces as Intelligent UI; fifteen months earlier Tyson had put a package name on npm for a related idea, and a calendar of who else had.

At 5:13pm on 9 October 2026 Tyson wrote to one of his agents: “looks like openai is going for intelligent ui, and interactive html. this related to generative ui and the older agent interface idea. seems obvious as a successor to chat.”1 dev sessions said 2026-10-09, session s:a96f2052, 17:13 AEDT.

Two days earlier OpenAI had announced GPT-6 with what it called Intelligent UI: “We’ve trained GPT‑6 to compose responses using text, visuals and interactive elements, choosing how they fit together based on your question.”2 OpenAI, “GPT-6 and Intelligent UI for everyone”, 2026-10-07, https://openai.com/index/gpt-6-for-everyone/.

At 5:55pm he sent a screenshot: “intelligent ui spotted in wild. good pattern. we should steal that.” A minute later: “let’s find a universal way of solving this better than chatgpt. we already have it with etc.”3 s:a96f2052, 17:55 and 17:56 AEDT; the tags are dev’s chat vocabulary. Those are tags his agents write into chat replies to put a choice, a permission or something to look at in front of him.

the older agent interface idea

On 30 June 2025 Tyson opened a new chat with Claude and pasted a status document for a project: a portfolio you talk to instead of read. The interface would change with the conversation.4 claude.ai chat 32c686de-158c-4af4-839b-b46b85a5afe0, turn 0, 2025-06-30 00:10 UTC, a pasted status document.

He liked the idea for a practical reason: “i hate shitdicking around trying to get my CV to fit on one page.” With a model in front, he could “dump as much detail as i want, actually the more the better,” and let it distil.5 Same chat, turn 2, 00:14 UTC. Twenty minutes into the chat he pasted a first UI: “this is what vercel v0 spat out.”6 Same chat, turn 8, 00:30 UTC.

The next evening the portfolio’s renderer had become a package name. He didn’t want it scoped under his username, which “makes it so inaccessible for other people especially if i want to create something akin to MCP for UI protocols.”7 claude.ai chat 0c8242f0-eb6a-4c39-ad72-e75bf58c2081, turn 2, 2025-07-01 09:21 UTC. At 9:54am UTC he pasted the registry’s reply, agentinterface@0.0.1 … published just now, and wrote: “Boom we are live.”8 Same chat, turn 44, 09:54 UTC, pasting npm view agentinterface. The package was 1.2 kB, and its README said “Coming soon.”9 Same chat, turn 40, the README he pasted before publishing: ”🚧 Under Development - Coming soon.” Size from turn 44: unpackedSize: 1.2 kB.

On 9 July he pasted a longer brief for it, the Agent Interface Protocol, into another chat. A section headed “Why This Is Novel” set it against the field: “Vercel AI SDK: 1:1 tool-to-component mapping.” Its own mechanism: “No mapping layer, no predetermined tool calls - just intelligent UI selection driven by conversational understanding.”10 claude.ai chat 882d00f2-6d21-4f62-be5a-80f63f0583b8, turn 46, 2025-07-09 18:26 UTC. A pasted brief; the comparison is its claim, unchecked.

The repository started on 10 July; within a week it had intent-based component selection. On 27 August a commit added the thing the brief had ruled out, an interpretation layer: a Python wrapper now hands an agent’s text and a list of available components to a second model, which returns a component tree as JSON for a React renderer to draw.11 ~/repos/archive/agent-interface, root d4be9c5 2025-07-10; python/src/agentinterface/ai.py l.125–137, shaper.py l.124–150, react/src/renderer.tsx l.28–83, as of HEAD; intent-based selection 587dbe2 2025-07-16; response shaping 6fbdb79 2025-08-27. The README: “LLM selects components. React renders. Infinite composition.”12 Same repo, README.md l.5. Version 1.0.0 went to npm on 22 October 2025.13 Same repo, edd1e7c, 2025-10-22, “v1.0.0: Production release”; npm registry, npm view agentinterface time.

the calendar

The brief compared itself to three products: Vercel’s AI SDK, LangChain and ChatGPT Plugins. Laid out by date, the neighbourhood was more crowded:14 Vendor primaries: vercel.com/blog/announcing-v0-generative-ui (2023-10-11); vercel.com/blog/ai-sdk-3-generative-ui (2024-03-01); anthropic.com/news/claude-3-5-sonnet (2024-06-21); copilotkit.ai/blog/copilotkit-v1-launch (2024-07-08); openai.com/index/introducing-apps-in-chatgpt (2025-10-06); github.com/MCP-UI-Org/mcp-ui (2025-05-13, repository date); research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt (2025-11-18); developers.googleblog.com/en/introducing-a2ui-an-open-project-for-agent-driven-interfaces (2025-12-15). Not exhaustive; scribe’s search found no pre-July-2025 protocol of the registry shape.

  • October 2023: Vercel announces v0, the tool that drew his first screen, under the heading “Generative UI”.
  • March 2024: Vercel’s AI SDK 3.0 streams React components chosen by a model’s tool calls, at runtime.
  • June 2024: Anthropic ships Artifacts, documents and small apps rendered beside the chat.
  • July 2024: CopilotKit 1.0 returns React components from agent actions.
  • May 2025: MCP-UI attaches interactive UI to agent tools, mostly as HTML.
  • July 2025: agentinterface 0.0.1, a placeholder.
  • October 2025: OpenAI’s Apps SDK puts developer-built app interfaces in ChatGPT; agentinterface 1.0.0 follows sixteen days later.
  • November 2025: Google’s Gemini generates custom interfaces per prompt.
  • December 2025: Google’s A2UI, an open format in which an agent composes from a catalog of components the host advertises.
  • October 2026: OpenAI’s Intelligent UI.

A2UI is the closest match to what the 2025 brief described: an agent choosing from a registry of components, through a protocol anyone can implement. It arrived five months after the first agentinterface commit and two months after its 1.0.0. The difference the brief claimed against Vercel, a model choosing freely instead of a tool mapped to a component, is a narrow one: Vercel’s own launch FAQ allows prompting a model for structured data instead of tool calls.15 vercel.com/blog/ai-sdk-3-generative-ui, FAQ: “What if my LLM doesn’t support tools or function calling?”

What the record supports is a package name on npm in July 2025, and a component protocol in the repository that month. “Intelligent UI” appears in his 2025 brief as a description and in OpenAI’s 2026 post as a product name.

the same week, again

On 12 October at 3:41am an agent in his lab committed a field note on the coordination tools that had shipped around him. Anthropic had announced a redesigned Claude Projects on 17 September: a coordinator conversation that routes work to worker threads, which report back and can call subagents. The note’s verdict on the shape: it “matches dev’s chief/lead → dispatched session → scout/advisor tree.”16 ~/lab 1bd2abc, docs/coordination/landscape-2026-10-12.md l.11, l.51; the announcement is claude.com/resources/articles/projects-redesigned. Dev is the tooling he runs his agents in. Its first session-launches-session commit is dated 3 September 2026, fourteen days before Anthropic’s announcement.17 ~/dev 98834a9a4, 2026-09-03 02:56 AEST, “launch_session tool and /handover succession”. Earlier multi-agent repos of his go back to 2025.

He had been up reading about the field already. At 3:30am, on a funded coordination startup: “openai and anthropic are already going to solve this / already have solved it.” At 3:32am, on Claude Projects: “i wonder if anthropic just converged exactly on dev.” Three minutes after the note landed: “might be the nail in the coffin for dev as a product. everyone is already converging and coordination is table stakes.”18 s:8fbf8148, 2026-10-12 03:30, 03:32 and 03:44 AEDT.

At 3:47am he sent scribe both cases, the interface and the projects: “just something to chew on.”19 s:ec5a4633, 03:47 AEDT.