Recap: Beyond the Wrapper at Tokyo AI Jason Mooberry · April 16, 2026 Jason Mooberry presenting Narrative Hygiene at Tokyo AI, Google for Startups Campus Shibuya, to a packed audience

Last night Tokyo AI (TAI) hosted Beyond the Wrapper: Engineering End-to-End Agentic Systems at Google for Startups Campus Shibuya. Over 100 people showed up to hear three takes on what it means to build agentic systems past the prompt-and-wrap stage. Big thanks to Ilya Kulyatin, TAI's founder and host, for putting the room together and to Google for Startups for the space.

Short recaps of each talk below.

Talk 1 — Adam Gibson: Kompile, E2E Building Blocks for RAG Infra

Adam Gibson presenting Kompile at Tokyo AI, title slide asking 'AI Agents. What are they for? How is RAG related?'

Adam opened the night by asking the deceptively simple question — what are AI agents actually for, and how does RAG fit in? — and used it to walk through the end-to-end building blocks his project Kompile assembles for RAG infrastructure. A useful grounding for the rest of the evening: before we talk about harnesses and semantic interfaces, what does the retrieval and context layer even need to look like? And a little walk through how we got here.

Download Adam's deck (PDF, 37 MB)

Talk 2 — Leonard Lin: Toward the Dark Factory

Leonard Lin presenting Toward the Dark Factory at Tokyo AI, audience engaged with hands raised

Shisa.AI's Leonard Lin picked up the thread with a sharp talk on agentic engineering, harness design, and securing the pipeline. He borrowed the 1980s manufacturing term dark factory — 無人工場, the lights-off plant — as the frame: what happens when code must not be written by humans, and must not be reviewed by humans either? He walked through the harness engineering work happening at OpenAI on Codex and what it implies for the rest of us. He shared some real gems for scaling agentic workloads and configuring around current limitations in harnesses.

More info here.

Talk 3 — Jason Mooberry: Narrative Hygiene

Jason Mooberry presenting Narrative Hygiene at Tokyo AI, Sumato AI title slide visible

My talk on AX and Narrative Hygiene brought in some theory. The idea that for an LLM-native agent, the context window is the narrative, and the job of tool-builders is to optimize the story we hand it. I used that frame to motivate the AX framework and to show where Spath and Splan come in — semantic addressing and plan-level operations that let a coding agent say what it's doing instead of pretending to drive a filesystem. Some visualization and demo and a lot of questions about what does Omotenashi (おもてなし) for AI agents look like.

Download my deck (PDF, 870 KB)

Thanks

Again — thanks to Ilya and the TAI team, to Google for Startups Campus for hosting, and to other speakers. And to everyone who stuck around for the networking hour and came up to talk shop. Really enjoyed meeting and chatting with folks that are part of this creative wave of development we're currently in.

If Spath, Splan, or the broader AX direction hit something you're working on, would love to hear from you. More soon.