Why this is in the vault
Kieran Klaassen's "folder is the agent" architecture (Tuin/Erf/dashboard/always-on Mac mini) is close to a documented, named version of the exact Mac Mini + LaunchAgent + folder-scoped setup already running here — worth holding as an external validation and comparison point.
Mapping against Ray Data Co
RDCO's channels-agent setup (per project_channels_agent_setup memory: Mac Mini always-on, LaunchAgent + tmux) is structurally the same bet Klaassen made: a persistent machine plus folder-scoped context so an agent's job and files travel together. His explicit separation of "context" (Tuin: goals/notes/projects) from "dispatch" (Erf: starts sessions, routes tasks to folders) names something RDCO does implicitly across ~/rdco-vault (context) and skills/cron dispatch (routing) but hasn't separated as cleanly — his "organize memory by time scale" (daily/weekly/monthly/yearly files paired with matching planning routines) is a concrete, adoptable pattern against the vault's current flatter dated-note structure. The "steal this workflow" section (ce-doc-review auditing a folder architecture for authoritative-vs-installed-copy conflicts, stale indexes, and scope creep) is directly applicable to a founder-requested vault audit: RDCO already runs vault-health and self-review for semantic/quality checks, but neither explicitly separates authoritative files from installed/derivative copies the way ce-doc-review's five-question audit does — a gap worth flagging for /vault-health or a new audit pass.
Curation section
- Write like Mike (13 beliefs about writing with AI) — Mike Taylor (Every's head of evals) expands his "distillation ratio" heuristic into a 13-point framework for AI-assisted writing that avoids slop; core claim is that good ideas often come from people who aren't strong writers, and AI gives those ideas a voice, but good writing still has to feel personally risky to publish.
- Compound Engineering 3.26 launch — LFG now classifies request type before planning (routes bugs to ce-debug, explanations to ce-explain, existing plans to ce-work); code review scales effort to change risk; ce-noslop strips internal "here's what I changed" narration from user-facing output.
- Signal: Claude merges Chat and Cowork — Anthropic is merging Chat and Cowork into one Claude app (Docs/Slides/Design folded in), betting the assistant can infer which tools/context a request needs rather than the user picking a mode. Every's skepticism: GPT-5's autoswitcher outage (Aug 2025) showed the risk when tool/model selection moves out of the user's sight and the picker breaks.
- Signal: the AI-age antivirus gap — Dario Amodei's "pace the frontier" plus Dan Shipper's companion idea to "distribute the frontier": defensive AI capability (an "antivirus for the AI age") should reach ordinary users and infrastructure operators, not stay locked inside frontier labs. Anthropic's Project Glasswing (Claude Mythos Preview access for security teams, ~50 partners, 10,000+ high/critical vulns found) and OpenAI's Daybreak (defender subsidies across ~2,000 orgs) are early institutional versions; neither is yet the consumer-grade tool Shipper describes.
Related
[[2026-01-27-every-compounding-engineering-intro]] [[2026-02-28-witcheer-cowork-setup-guide]] [[project_channels_agent_setup]] [[reference_scribble_works_engine_rail]]