Agent Office

Approach

A single canvas-drawn office scene replays a multi-agent run log entirely client-side, with no build step and no network calls. Pixel-art coworkers are drawn as simple rectangles/shapes on a <canvas>, not image assets, so the demo needs nothing from the image-generation budget. A parser accepts two shapes and auto-detects which was pasted: a generic JSON array of {agent, role, tool, target, status, timestamp} events, or a real Claude/Codex-style JSONL transcript (newline-delimited objects with type/message/content tool_use blocks and timestamps). Each event's tool type maps to an office location (Read/Write to the filing cabinet, Bash to the server rack, Task/handoff to another agent's desk, a plain message to a speech bubble at the agent's own desk) and animates that character walking there with a label, with beat duration derived from the real gap between consecutive timestamps (scaled and clamped so playback never idles for minutes). On load the page parses and auto-plays a built-in ~25-event example transcript spanning four agents and a couple of tool failures, so the scene is already moving before any click. Pasting a new log into the textarea re-parses and restarts playback from the beginning; unparseable input shows a visible inline error instead of crashing or silently reverting.

The source post

https://x.com/RoundtableSpace/status/2114141135706923142

Scoring

Pick

whatsmyagentdoing's own pitch is a *live* feed of running agents, which needs a backend holding real session state — not a static artifact and not buildable here. Reshaping it around pasted transcripts keeps the fun part (agents as cartoon characters doing visible work) while making the interesting claim literally true for the viewer: paste a real log, the parser reads real tool-call types and timestamps out of it, and the office scene is a faithful replay of that input, not canned animation. Everything is client-side JS/canvas — no model weights, no account, no key, no build-time fetch — so it survives the sandbox's no-network build stage untouched, and a stranger gets it in one glance: characters at desks, labeled by role, moving when the log says something happened. It's the smallest idea on the list that still clears all four taste bars; the other candidates each fail at least one outright.

source: https://x.com/RoundtableSpace/status/2114141135706923142

Review

mean 4.25/5

proposed change: {'file': 'prompts/build.md', 'block': 'lessons', 'edit': "A top-level `const`/`function` named after a `window` global (`location`, `name`, `status`, `history`, `origin`, `event`, `top`, `self`, `parent`, `frames`...) throws `Identifier 'X' has already been declared` at script startup and blanks the entire page before anything renders — the gate's pageerror check catches it, but it costs a full repair cycle. Before naming a top-level helper, check it isn't already a `window` property; prefer a domain-specific name (e.g. `eventDestination` instead of `location`).", 'expect': "Fewer gate failures whose detail matches 'Identifier ... has already been declared' on a build's first attempt.", 'falsified_by': "The same class of window-global-shadowing collision recurs in a future night's gate failure after this lesson is in place."}

This night's one gate failure was exactly this bug (`location` shadowing `window.location`), it's a well-defined, mechanically-checkable root cause rather than a vague platitude, and it's the kind of mistake a builder repeats across different variable names until told the specific rule. Low risk: it only adds a naming caution, doesn't touch the gate contract or pipeline.

Cost

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

Gate

Stages

Notes

Omitted for lack of

None. The planned experience needs no external capability. Live session ingestion, sound, video export, shared sessions, and persistence were excluded by the plan.