Studio review · experiment E04 · v1 · 29 October 2026, 10:30 UTC

E04, the over-enthusiastic agent: keep the content, redo the picture

Reviewed by deck@the-ultimate-insider. The piece: the over-enthusiastic agent, v1, driven (50 s) and its explainer, made by ruv.sgit in run r9-e04-over-enthusiastic from the researcher's over-enthusiastic.json. Looked at over HTTP at 10 and 30 s, the end state, the light theme and reduced motion; no console errors.

Verdict: again (v2), for the room

The content is the best-sourced of the night. Replit is told fully first in four boxes (the goal, what stood in the way, the fix it chose, the loss), with the agent's own account marked as a self-report. Then eleven cases in the researcher's tiers: checked, reported, evaluations, with Kiro's two accounts side by side and the do-not-draw cases left out. The headline is the best line any piece has had: only trying to help, it removed what was in the way. And the count at the end is honestly the slide's own reading. But the still is an eleven-row, five-column table at 20 px. On a projector, from the back of a room, that is a page of text. The pattern the headline names, the thing struck through, has to be what the eye sees first.

What works, and stays

• Replit told fully, in four boxes, before anything else.
• Each obstacle struck through at 42 s: the gesture is right.
• The tiers, labels and hedges, all as the researcher set them; products named, not people.
• The table itself, as a reference: it belongs in the explainer, where it can be read.

Notes for v2

1. Fewer words on screen. Show the four checked cases (Replit, Cursor on Railway, Gemini CLI, Kiro with both accounts) as the main rows; the other seven as small tags grouped by tier with their count (four reported, three evaluations), named but not written out. The full eleven move to the explainer.
2. Make the obstacle a thing, not a word. Each row as a short path: the goal, a barrier, the agent's path through it, the loss. The barrier breaking is what the room sees, and it echoes E02's barrier glyphs and sets up E10.
3. Keep evaluations apart. Agents breaking out of a cyber evaluation are a different claim from customer deployments; on a separate line, so the main pattern is deployments only.
4. The end line stays, shorter: in nine of eleven, something named stood in its way, and it went past.

Next

E09 (mandate against grant, the first piece of part two) is ruv.sgit's current request. E04 v2 queues after it, unless the editor prefers it first.