Miadi Review

Engineering Review: Episode 120 Screenwalk to Navigable Research Architecture

Oct 15, 2026 · model: manual-import · channel: JGWill

Watch source

In the chronicle

Engineering Review: Episode 120 Screenwalk to Navigable Research Architecture

Source: ep120-screenwalk-261012-mia-branches-marks-jev-ep339-voice — JGWill https://youtu.be/Us0oZL_EObY

Summary

William walks through Episode 120 of The Chronicle and shows a fork-and-harvest way of working that links narrative storytelling to technical work a person can inspect. Using tmux terminal sessions, Claude Code, and the Gaia voice companion, the operator directs sub-agents on three branches in parallel: plot threads, annotation marks, and Jev scoring. The workflow rests on relational consent before dispatch and on links to Langfuse traces and GitHub issue jgwill/Miadi#754. The video presents these as a navigable research interface deployed on Vercel. Figures and timecodes below are as summarized from the video, not checked.

Academic Fields

Human–AI Interaction (Human–Computer Interaction)

Studies how people direct, correct and understand AI systems. The video touches it where the operator steers agents by voice from a phone and Mia answers by speech. Amershi et al. give 18 design guidelines, tested by 49 practitioners on 20 products. They guide design; they do not show one setup works for others.

Multi-Agent Systems (Computer Science)

Studies how several agents divide work and share results. Fork-and-harvest is an orchestrator-worker pattern: a trunk forks branches, then collects their reports. Anthropic reported about 15 times a chat's tokens and weaker fit where agents depend on each other. Cemri et al. catalogue 14 failure modes, so one demonstration shows possibility, not reliability.

Provenance and Research Data Management (Information Science)

Studies records of origin: who did what with which inputs, so others can judge a result. The video's doorways from a scene to a GitHub issue, Langfuse trace and saved prompt are provenance links, close to W3C PROV's entities, activities and agents. Provenance shows origin, not correctness, and FAIR asks for persistent identifiers.

Human Oversight in Software Process (Software Engineering)

Studies where people stay in the loop of automated work. The video separates marking a session from dispatching work; Parasuraman, Sheridan and Wickens describe automation by stage and level. The limit: reviewers of automated output grow complacent and over-trust it (Parasuraman and Manzey), so an approval step alone does not show oversight works (Green).

Indigenous Research Methodology (Research Is Ceremony, Wilson 2008)

Wilson, an Opaskwayak Cree scholar, describes research as ceremony: accountability is owed to "all our relations", from choosing topics to presenting findings. The video borrows relational accountability for consent before dispatch. Wilson wrote about research among people, not software agents, so this use is the video's extension, for Indigenous scholars and rights-holders to judge.

Narratology and Transmedia Storytelling (Literary and Media Studies)

Narratology separates story, the events, from discourse, how they are told (Chatman). Transmedia spreads a story across media, each doing what it does best (Jenkins). The video keeps plot threads apart from tmux branches and links scenes to evidence. Jenkins asks each entry to stand alone, so linked issues and traces document, not narrate.

Aspects

Timecodes in this review were generated by the service from the video and were not checked against the footage. YouTube's page gives the video's runtime as 57:01, and none of the timecodes runs past it.

Trunk and Execution Threads Separation (17:03-17:47)

Plot threads govern narrative arcs while execution threads track terminal tmux sessions managing sub-agent tasks [17:03-17:47]. The Episode 120 briefs call a tmux session a "branch" and use "plot thread" only for a concern followed from setup to resolution. The human-led script acts as the trunk anchor before sub-agents fork into parallel execution states in separate sessions, each worked independently until its results are harvested.

Fork-and-Harvest Multi-Agent Coordination (17:48-18:56)

The pipeline coordinates sub-agents across three specialized branches: plot terminology modeling, annotating marks via Plannotator, and scoring via Jev [17:48-18:56]. Each branch operates concurrently in its own tmux session before its structured state models, annotation marks, and scores are harvested back into the main trunk.

Relational Consent Boundaries in Mark Dispatch (18:57-20:00)

Marking a session records annotations on an agent's output; dispatching work requires an explicit, negotiated delivery agreement [18:57-20:00]. The video grounds this in Shawn Wilson's relational methodology (Research Is Ceremony, 2008): decisions and dispatching call for shared consent, and relational accountability stays inside the circle.

Doorways to Inspectable Evidence (20:01-21:06)

A navigable research architecture links the narrative to the artifacts behind it through clickable doorways [20:01-21:06]. From a screenwalk scene a reader can open GitHub issue jgwill/Miadi#754 (open when checked on 2026-10-14), a Langfuse trace holding 13 observation records (as summarized from the video, not checked), and a reusable prompt, without the earlier chat. A link lets a person check; it does not show that what is linked is correct.

Auditory Multimodal Steering via Gaia Companion (32:20-44:50)

The operator uses the Gaia web audio interface on mobile to issue voice takes directly into running terminal sessions [32:20-35:00, 42:00-44:50]. Agent Mia ingests speech, steers branch execution in tmux, and replies through speech synthesis with fallback to on-device TTS.

End-results and Core Aspect

Core Aspect — Navigable Research Architecture with Fork-and-Harvest Scaffolding (15:15-16:54)

A research interface that turns raw screenwalk captures into navigable pages linked to their evidence [15:15-16:54, 53:50-55:00]. It links narrative to the technical artifacts behind it, so human operators and sub-agents can branch into parallel threads and bring evidence back into the chronicle, where a person can inspect it.

Results

  • Interface Analysis Page (53:50-55:00): The video shows miadi-factory-interface-analysis.vercel.app, which presents the episode narrative with a State Loom view, a Langfuse evidence room, and links to GitHub issue jgwill/Miadi#754 [53:50-55:00]. The page answered HTTP 200 on 2026-10-14, and the Episode 120 manifest lists it as a prototype made with v0 (issue jgwill/Miadi#757). What the State Loom and Langfuse views contain is as summarized from the video, not checked.
  • Voice-to-Tmux Steering Across Panes (42:00-53:10): Voice takes sent from a phone through Gaia are followed by sub-agent activity in several tmux panes: task management, logged outputs, and updated branch agreements [42:00-45:30, 51:00-53:10]. A demo shows this happening once; it gives no measure of how often it works or how long it takes.
  • Langfuse Trace Ingestion and Reusable Prompt (20:30-21:06): The video describes 13 observation records written to a Langfuse trace (as summarized, not checked) and a reusable prompt saved so a later agent can read the trace without the earlier conversation [20:30-21:06, 54:30-55:00]. The Episode 120 manifest names that prompt miadi-ep120-fork-voice-layer-ingestion-261009, version 2, label draft. The review's shorter name omits the date suffix.

Sources