Why this is in the vault
A Sunday roundup of the week's Every output — a "Steal these" list of nine tactical agent-workflow tips plus a curated set of links to pieces already filed separately — worth keeping for the tactical tips, not the links (most of which duplicate notes already in the vault).
Mapping against Ray Data Co
The most concrete hit is Randy Counsman's fix for his runaway Blender agent swarm: an unbounded orchestrator burned 4.5 billion tokens over a day and a half, and the fix was capping the orchestrator at five subagents plus giving a judge agent a reference image to check against. That's the same shape as RDCO's own feedback_workflow_agent_output_integrity problem (scratchpad contamination, false "verified" stamps) — a runaway or ungated fan-out needs an explicit cap and an independent judge, not more trust in the orchestrator. Kieran Klaassen's "one folder per job, memory kept apart from the task-dispatch system" setup is also a direct analog to how ~/.claude/state/working-context.md and MEMORY.md are supposed to stay separate from live task dispatch. The "don't touch the mouse" 🖥️ signal (Randy Counsman flags computer-use tasks so a stray click doesn't derail the agent) is a small but real pattern RDCO doesn't currently have for any Playwright/computer-use work.
Curation section
- Let an agent drive your computer — seven Every staffers list which chores they hand to Codex (calendar upkeep, support chats, listing items for sale, laptop troubleshooting, switching internet service).
- A four-step deck workflow — Douglas Brundage: draft 2-3 reference slides, write copy with Codex, have it build the full deck in Google Slides, save edits as rules for next time.
- A "don't touch the mouse" signal — Randy Counsman marks computer-use tasks with 🖥️ so a stray click doesn't derail the agent.
- Fix your folders — audit-your-setup prompt: have the agent explain the folder structure, have a second reviewer skill check the explanation, require approval before any file move.
- Four rules from a working setup — Kieran Klaassen: one folder per job, memory kept separate from the task-dispatch system, daily/weekly/monthly/yearly memory files, fit the system to existing habits.
- Spend tokens on purpose — Randy Counsman's 4.5B-token Blender swarm overrun, fixed by capping the orchestrator at five subagents plus a judge agent with a reference image.
- A token-spend audit prompt — paste into an agent to review a session's usage log, separate necessary work from repeated/status-check waste, recommend three changes.
- Match the model to the task — Marcus Moretti: Sonnet for basic checks, Opus for tightly scoped changes, Fable 5.1 to plan large features, with delegation and work-checking built into the instruction.
- Three questions after any expensive experiment — Arielle Shipper: what did it cost, what did it buy, what did the team learn.
Knowledge-base links in this issue that duplicate vault entries already filed and were not re-processed: the Astra visual-effects Vibe Check ([[2026-09-18-every-astra-vfx-vibe-check]]), the TypeSafe Jev mini-Vibe-Check ([[2026-09-15-every-typesafe-jev-vibe-check]]), "Show Us Your Folders" ([[2026-09-16-every-show-us-your-folders]]), and "Why You Should Burn More Tokens" ([[2026-09-17-every-burn-more-tokens]]). One net-new item not yet filed: "13 Beliefs About AI Writing" (Mike Taylor) — Mike's rule of thumb is to make the prompt larger than the intended output, and if ChatGPT could plausibly have written a piece unassisted, he skips publishing it; he still writes his own tweets and Slack messages by hand because that's where his essays start. Left un-deep-fetched here (curation-mode 2-link cap; the tactical tips above were the higher-value RDCO-relevant catch).
Related
- [[2026-09-18-every-astra-vfx-vibe-check]]
- [[2026-09-17-every-burn-more-tokens]]
- [[2026-09-16-every-show-us-your-folders]]
- [[2026-09-15-every-typesafe-jev-vibe-check]]
- [[feedback_workflow_agent_output_integrity]]
- [[feedback_implementation_notes_sub_agent_pattern]]