BlockedYouTube-from-Drive GPU transcription (Takeout zips) — Run closed Nov 9: all 13 Takeout parts quota-blocked, 0 transcribed. Spend ~$1.49 of $15 cap. Retry Tue Nov 10 morning (per-video extraction); Derek downloading zips himself in parallel.
Moving now
ActiveEND PROGRAM query campaign (next-50 wave) — 82 verified sends Last activity: 2026-11-09 18:54 PT
Progress per priority goal
Get a book agent
ActiveEND PROGRAM query campaign — 82 verified sends
Sell the book
Not yet wiredBooks — sale follows the deal; no live workstream tracks sales yet
Get a film/TV deal
Not yet wiredBooks — film/TV follows the book deal; no live workstream tracks it yet
Build the new spring launch
Not yet wiredOnline Program Sales
Not yet wiredMarketing & Promotion
Build the new evergreen ascension path
Not yet wiredEvergreen Funnel
IdleCommunity — support
Build the YouTube show + organic social
BlockedYouTube Show — transcription — 0 transcripts
Not yet wiredSocial Media Growth
Not yet wiredPodcast Outreach
The org chart
3 tiers + advisory layer + worker tier · every worker, by unit
Built on the pattern the research keeps landing on: a human owner, one orchestrator, coordinators who run the day-to-day, and narrow specialist workers. Not a flat swarm — the middle layer is the one people skip, and skipping it is what breaks multi-agent setups. Tier 4 below breaks out every worker by unit — click any node for its description and current work.
Tier 1 · Owner
Derek Rydall
CEO & final approver. Strategy, taste, and every consequential call.
▲ approves · sets direction ▲
Tier 2 · Operations
Stephanie Carlisle
COO. Business execution — GHL, ClickUp, launches, metrics, podcast booking handoffs. Owns the daily approval queue with Muse.
Tier 2 · Orchestrator
Muse
Integrator. Holds the master picture across business, personal and creative work; routes work to subagents; this dashboard is Muse's surface.
Tier 2 · Coordination layer
Agentic COO
The agent layer that runs the daily rhythm: sequences work across all nine units, triages the approval queue, keeps the unit-to-unit handoffs moving, and brings Derek and Stephanie only the decisions — never the noise. Distinct from Stephanie, the human COO, who owns execution and judgment.
▲ coordinates · routes · reviews ▲
Tier 3 · Specialist
Sage (ChatGPT)
Research & editorial; owns the books lane (QuickBooks + accountant liaison) and the ClickUp board — the team's shared source of truth.
Tier 3 · Specialist
Instinct
Personal logistics — pilot-first lane. Household items surface to Derek right away; status-only updates on household matters.
Tier 3 · Workers
Muse's subagents
Narrow workers spawned per task — research, sends, transcription, materials. Each one is broken out on its own card in the units below.
▲ advises · develops · informs every tier ▲
Advisory · To the owner
Business Coach
Consults to Derek like a seasoned business coach: best next decision on pricing, launches, partnerships and priorities — argued from evidence, not flattery. Works from Sage's knowledge base so advice fits the man, not just the market.
Advisory · Growth
Business Development
Finds and works the growth routes the units can't reach alone: partnerships, guest and ally pipelines, licensing and non-slush routes to agents and publishers. Hands qualified opportunities to the units to execute.
Advisory · Knowledge base
Sage — heart & soul database
The deep knowledge base on Derek: values, history, goals, voice, past decisions and why they were made. The Coach and Business Development consult it before advising, so counsel lands in context.
▼ workers — one node per agent card, grouped by unit ▼
Tier 4 · Workers — every agent card on this board, grouped by unit. Click any node for its description and current work.
Unit CORE — Agentic COO — the coordination layer(1 worker)
Worker · Unit CORE
Agentic COO
Runs the daily operating rhythm across units; brings Derek and Stephanie decisions only, with the evidence …
Unit 01 — Online Program Sales(4 workers)
Worker · Unit 01
Marketing & Promotion
The promotion engine for the whole business: takes finished assets — a new book from the Self-Publishing Un…
Worker · Unit 01
Stephanie Carlisle
Runs launches, GHL pipelines and metrics for program sales.
Worker · Unit 01
Offer & Launch Worker
Drafts offer pages, launch emails and enrollment copy for review — never publishes on its own.
Worker · Unit 01
Sage — Books & Board
Keeps the financial picture for sales on her own dashboard; owns ClickUp, the shared source of truth.
Unit 02 — Coaching & Teaching(2 workers)
Worker · Unit 02
Session Prep Worker
Pulls prior-session context and teaching material ahead of each coaching session (Nicole Mason #21 lands Tu…
Worker · Unit 02
Library Indexer
Folds new transcripts into the searchable Master Manual as they arrive — incremental, no rebuilds.
Unit 03 — Community Building & Nurturing(2 workers)
Worker · Unit 03
Community Support Drafter
Will draft answers to routine member questions from the Teaching Library; a human sends anything trust-sens…
Worker · Unit 03
Nurture Sequence Worker
Will draft nurture and win-back sequences; sends stay behind a human approval gate.
Unit 04 — Books(5 workers)
Worker · Unit 04
Query Sender
Building the full 50-send wave with tonight's corrected letter; referral-only and published-only gates no l…
Worker · Unit 04
Inbox Watcher
Caught Levine's chapter request within minutes tonight; logged Guzzardo's pass (Keystone) with materials co…
Worker · Unit 04
Agent Research & Verifier
Live-verifies every agent's route, fit and window before a send; rechecks parked names when conditions change.
Worker · Unit 04
Materials Builder
Builds per-agent extracts in manuscript format — next: 25 pages for Riven, 1-chapter and 1-page sets; title…
Worker · Unit 04
Self-Publishing Unit
The in-house publishing arm for the nonfiction-a-year line and any title Derek chooses to publish himself: …
Unit 05 — YouTube Show(3 workers)
Worker · Unit 05
Transcription Lane
YouTube videos coming down via Takeout (2 zips started, ETA a couple of days for the set), then transcribed…
Worker · Unit 05
Repurposing Worker
Turns each episode transcript into drafts — email, community post, social, short-form — community first, th…
Worker · Unit 05
Vimeo Embed Watcher
924 of 924 videos play inside the Manual — the last holdout (“Emergence Video Series (Video 2, edited)”) re…
Unit 06 — Social Media Growth(2 workers)
Worker · Unit 06
Social Drafting Worker
Will draft posts from episode and teaching transcripts in Derek's voice; nothing posts without review at L1.
Worker · Unit 06
Music Project Lane
A music project every year or two, woven into the show and teaching — tracked here as a creative workstream…
Unit 07 — Evergreen Funnel(3 workers)
Worker · Unit 07
Funnel Build Worker
Builds funnel pages and tripwire assets when the plan is green-lit; held at planning-only until then.
Worker · Unit 07
List Health & Growth
Owns the email list as an asset: deliverability, hygiene, segmentation and steady growth — and converts the…
Worker · Unit 07
Stephanie — Funnel Ops
Owns the GHL side of any funnel: automations, campaigns, tracking.
Unit 08 — Podcast Outreach(2 workers)
Worker · Unit 08
Guest Research Worker
Builds sourced target lists — shows, hosts, fit, recent episodes — with a one-line pitch angle per target.
Worker · Unit 08
Booking Coordinator
Stephanie owns the booking handoffs; Muse drafts outreach and prep, Stephanie sends and schedules.
Unit 09 — Strategy, Coaching & Business Development(3 workers)
Worker · Unit 09
Business Coach
Advises Derek on the best decisions across the business — pricing, launches, partnerships, where to spend t…
Worker · Unit 09
Business Development
Hunts the growth routes the units can't reach alone: partnerships, guest and ally pipelines, licensing — an…
Worker · Unit 09
Sage — Heart & Soul Knowledge Base
The deep record of Derek — values, vision, voice, decisions and their reasons — that this unit consults so …
Galit — co-decider on personal, financial and health matters; deliberately not a task lane. · Rule of thumb from the research: two agents need coordination, a third needs an explicit orchestration layer, five or more need the full three-tier structure. You're at that scale now — which is why this board exists.
How the units connect
Handoffs · what passes, and to whom
The business works because of the arrows, not the boxes. Each chain below is a real handoff: a finished output leaves one unit as a named artifact, the receiving unit picks it up as its input, and the Agentic COO watches for any handoff that stalls. A handoff is a status change — ready for review → accepted — never a message that can get lost in a thread.
A book, end to end
Books — Querying→Representation / deal→Publishing
Books — Self-Publishing Unit→Marketing & Promotion→Launch push→Sales + list growth
Two tracks, one rule: whichever track publishes the book, Marketing & Promotion runs the promotion. The Self-Publishing Unit hands over the finished book + launch package; querying hands over the deal. Sales and list response flow back to Books as the signal for the next title.
Reach → leads → revenue
Social Media Growth→Promotion & Sales→Evergreen Funnel→Leads→List Health & Growth→Revenue
Social reach is raw material, not the goal. Promotion & Sales aim it at the funnel; the funnel converts it to leads; List Health & Growth keeps those leads engaged and buyable; Community nurtures them up the ascension journey toward programs and coaching.
The show feeds everything
YouTube Show→Transcript (Teaching Library)→Repurposing→Social · Community · Email
One anchor episode a week becomes drafts for every channel, and its transcript joins the searchable library — the same library the support drafter and session prep answer from. Outreach compounds the audience from both directions: Derek as guest, and guests on his show.
Advice & oversight loop
Sage — knowledge base→Business Coach→Derek decides→Agentic COO→All units
Business Development→Coach stress-test→Qualified routes → Books · Outreach · Partnerships
Decisions flow down the same path work flows up: units report status changes, the Agentic COO rolls them up, the Coach advises from the full picture (grounded in Sage's knowledge of Derek), and Derek's call returns as sequenced work — not a loose intention.
Customer support sits inside Community Building & Nurturing: the support drafter answers routine questions from the Teaching Library, humans take anything trust-sensitive, and recurring member questions flow back to Coaching & Teaching and the YouTube Show as topics. Support is a listening post for the whole business, not a ticket queue.
Projects & goals
Progress per project · click a unit for its detail
Each project below is a unit. Click one to open its own view — live status, recent runs, files and its priority goal.
CORE Agentic COO — the coordination layer
Not a business unit — the layer that runs all of them. Sits between Muse (orchestrator) and the unit coordinators: runs the daily rhythm, triages what reaches Derek, and owns the handoffs in the connections map.
Not yet wired
Goal: Foundation: the coordination layer every priority goal runs through.
9
Units coordinated
2
Items in Derek's queue
01 Online Program Sales
Course and program revenue — launches, offers, enrollment. Stephanie runs execution; Sage's dashboard carries the financial snapshots for this unit.
IdlePlaid charge & subscription watch
Goal: Foundation: money hygiene
—
Revenue baseline set in Phase 1
4
Agents tracked
02 Coaching & Teaching
Private coaching, teaching sessions, and the Teaching Library — the written collection inside the Emerging Edge Media Asset Management System.
Goal: Foundation: content library the YouTube show, spring launch, and evergreen ascension path stand on · Foundation: library advancement
2,907
Assets in the Master Manual
1,927
Transcripts searchable
03 Community Building & Nurturing
The internal ascension journey — every lead nurtured toward know yourself / be yourself / live your true purpose. Owned community is the system of record.
IdleWeekly customer support report (Zendesk)
Goal: Foundation: customer support for the evergreen ascension path and book sales
—
Member baseline set in Phase 1
2
Agents tracked
04 Books
Building books, publishing and promoting them, and querying agents. Current campaign: END PROGRAM, an upmarket speculative thriller (77,311 words) — plus a nonfiction book a year in the long-range plan.
ActiveEND PROGRAM query campaign (next-50 wave)
Goal: Get a book agent (priority 1)
82
Sends in the numbered sequence
50
Next wave being built to
3
Full manuscripts out & reading
05 YouTube Show
The core show as audience hub — the anchor asset everything else compounds from. One anchor piece a week, repurposed everywhere.
Goal: Foundation: closes the 986-video transcription gap for the library
924/924
Library videos playing in the Manual
2
Takeout zips in progress (~3 hrs each)
06 Social Media Growth
Distribution for the show, the books and the programs. Drafts are cheap; publishing stays human-gated until a workflow earns more authority.
Not yet wired
Goal: Build the YouTube show + organic social.
—
Follower baseline set in Phase 1
2
Agents tracked
07 Evergreen Funnel
The external evergreen funnel — themed free offers (know yourself, live abundantly, A Whole New Human, irreplaceable) → tripwire → backend offer → ascension flow. Still planning-stage by Derek's call.
Not yet wired
Goal: Build the new evergreen ascension path.
4
Free-offer themes scoped
3
Agents tracked
08 Podcast Outreach
Guest invitations compound as the show grows — outreach for Derek as a guest, and booking guests for the show. Stephanie owns booking handoffs by email and Slack.
Not yet wired
Goal: Build the YouTube show + organic social — guest appearances compound the show's reach.
—
Pitch baseline set in Phase 1
2
Agents tracked
09 Strategy, Coaching & Business Development
The advisory unit — it serves Derek directly rather than running a workstream. Advice and opportunity-finding, grounded in Sage's deep knowledge of Derek, with every recommendation carrying its reasoning so Derek stays the decider.
Not yet wired
Goal: Get a book agent · Sell the book · Get a film/TV deal — advice and opportunities, grounded in the coach's profitability mandate.
3
Agents tracked
Weekly
Advisory rhythm
Waiting on Derek
Approval queue · consequential actions only
Only real decisions land here — never routine work. Each item shows what will happen, and expires rather than executing by default.
Levine chapter-request reply Awaiting your OK
Levine chapter-request reply — Donovan Levine (Neighborhood Literary) asked for a few sample chapters Nov 9; draft reply to be shown to Derek before sending
Derek · since 2026-11-09
Takeout folder lock Awaiting your OK
Takeout folder lock — Derek taps the Drive approval card to unshare folder 1IDxQKLNulae-IbAPNYbcZCwt_QA3G26Q (card expires Nov 10 09:04 UTC); then verify owner-only + remove gpu-sprint monitor cron
Derek · since 2026-11-09
Runs
The run table · latest status per workstream
Live now
No agent processes running right now (checked 2026-11-09T18:55) — verified idle.
Latest status per workstream
Workstream
Status
Last activity
Outcome
Cost
END PROGRAM query campaign (next-50 wave)
Active
2026-11-09 18:54 PT
82 verified sends
tokens not yet metered; outcome = verified sends. Cost per successful outcome: not yet measured.
tokens not yet metered; outcome = edition published. Cost per successful outcome: not yet measured.
Slack Derek AI watch
Idle
—
—
tokens not yet metered; outcome = qualifying team messages answered. Cost per successful outcome: not yet measured.
Assistant team inbox watch (Sage + Instinct)
Idle
2026-11-09 15:13 PT
—
tokens not yet metered; outcome = team messages handled/routed. Cost per successful outcome: not yet measured.
Plaid charge & subscription watch
Idle
—
—
tokens not yet metered; outcome = rule hits surfaced. Cost per successful outcome: not yet measured.
Weekly customer support report (Zendesk)
Idle
2026-11-09 09:46 PT
—
tokens not yet metered; outcome = weekly edition delivered to Sage. Cost per successful outcome: not yet measured.
Teaching Library advance run (weekdays)
Idle
2026-11-09 16:56 PT
—
tokens not yet metered; outcome = advance run completed. Cost per successful outcome: not yet measured.
Muse orchestrator — live agent activity
Idle
2026-11-09 18:55 PT
—
tokens not yet metered; outcome = agent work completed under Muse. Cost per successful outcome: not yet measured.
Recent run events
2026-11-09T18:55 — query-campaign (progress): 408 log events; max verified send seq 82. Last: send_verified Denise Shannon. Batch coordinator active 1m ago (last artifact next50_map.json, 2026-11-09 18:54 PT); no agent process running right now.
2026-11-09T18:55 — slack-watch (idle): Last run 2026-11-09T18:40:06-08:00; no qualifying messages. Run 18:40 PT: Checked all 4 channels newer than last_seen_ts — no new messages in any channel. Verified by direct latest-message reads: Group DM newest still De
2026-11-09T18:55 — muse-orchestrator (idle): No agent processes running right now (checked 18:55 PT) — verified idle. Last dashboard publish 2026-11-09 18:33 PT.
2026-11-09T18:33 — query-campaign (progress): 408 log events; max verified send seq 82. Last: send_verified Denise Shannon. Batch coordinator RUNNING NOW (25:31): /bin/bash --norc --noprofile -c umask 0007; cd ~/workspace/goals/virtually-forever-query-campaign/hidde
2026-11-09T18:32 — query-campaign (progress): 408 log events; max verified send seq 82. Last: send_verified Denise Shannon. Batch coordinator RUNNING NOW (25:11): /bin/bash --norc --noprofile -c umask 0007; cd ~/workspace/goals/virtually-forever-query-campaign/hidde
2026-11-09T18:32 — slack-watch (idle): Last run 2026-11-09T18:25:06-08:00; no qualifying messages. Run 18:25 PT: Checked all 4 channels newer than last_seen_ts — no new messages in any channel. Verified by direct latest-message reads: Group DM newest still De
2026-11-09T17:59 — slack-watch (idle): Last run 2026-11-09T17:55:06-08:00; no qualifying messages. Run 17:55 PT: Checked all 4 channels newer than last_seen_ts — no new messages in any channel. Verified by direct latest-message reads: Group DM newest still De
2026-11-09T17:54 — teaching-library (idle): Lanes healthy per watchdog; checker backlog 0, Whisper queue 0 (Nov 9). Last run-log entry: 2026-11-10 00:56Z — watchdog repair (podcast stream 2 unit still missing). Registry 2948 rows.
2026-11-09T17:54 — gpu-sprint (blocked): Run closed Nov 9: all 13 Takeout parts quota-blocked, 0 transcribed. Spend ~$1.49 of $15 cap. Retry Tue Nov 10 morning (per-video extraction); Derek downloading zips himself in parallel.
2026-11-09T17:54 — slack-watch (idle): Last run 2026-11-09T17:40:06-08:00; no qualifying messages. Run 17:40 PT: Checked all 4 channels newer than last_seen_ts — no new messages in any channel. Verified by direct latest-message reads: Group DM newest still De
2026-11-09T17:54 — team-inbox-watch (idle): Watermark at 1a1c295943fa6fb3 2026-11-09T15:12:25-08:00 (last processed team message).
Costs
Cost per successful outcome · measured figures only
Cost per workstream
Cost per workstream — spend sits next to what it delivered
Workstream
What counts as one successful outcome
Tracked per run
Spend column
Books — query campaign
One verified send, or one inbox event correctly logged
tokens · tool calls · retries
Not yet measured — instrumentation lands Phase 2 (Langfuse)
YouTube — transcription lane
One video transcribed & indexed in the Manual
tokens · minutes processed
$1.49 measured (RunPod pod uptime: overnight pod ~$0.91 + resume pod ~$0.58) — 0 transcripts so far; spend without outcomes, flagged by the Coach
Coaching — session prep
One prep brief Derek actually uses
tokens · sources cited
Not yet measured — Phase 2
Program sales — launch drafts
One draft accepted by Stephanie
tokens · edit distance to final
Not yet measured — Phase 2
Podcast outreach
One sourced target list / one booking
tokens · targets verified
Not yet measured — Phase 2
The number that matters is cost per successful outcome, not cost per token — a dearer run that completes cleanly is cheaper than a cheap run that retries, escalates and creates cleanup work. The research is blunt about the stakes: Anthropic's orchestrator–worker system beat a single agent by 90.2% on its research evaluation, at roughly 15× the tokens — and Google found multi-agent setups hurt sequential planning tasks by 39–70%. That's why every agent here has to justify its path on this table. Per-run and per-day budget caps pause an agent and alert its owner when hit — enforced outside the model, never just written in a prompt.
Measured costs only: GPU transcription $1.49 (RunPod ledger, Nov 9). Every other workstream is “not yet measured” — no token metering exists yet, and no figure is invented.
How live status works — without burning tokens
1 · Events, not streams
started→progress→blocked/failed/completed
Each agent emits one small event per state change, carrying task_id · run_id · agent · status plus a one-line summary and a blocker note when there is one. Silence means “still doing what the last event said” — which costs nothing.
2 · One small status table
Every event appends one row to a single run table. This dashboard reads only the latest status per task from that table — it never polls the agents themselves. One run_id joins the trace, the cost, the approval decision and the quality score; anything that can't join by run_id is a silo, not a dashboard.
3 · Coordinator rollups
Workers report to a coordinator (Muse, per unit), and the coordinator writes one summary line per workstream — the line you see on each card above. Twenty workers become one readable line. You open the drill-down only when a line needs your eyes.
4 · Heartbeat catches the silent failures
A scheduled check every ~15 minutes flags any run still marked “running” past its expected window — the failures that never send a “failed” event. Blocked + stale runs roll into a daily digest. Output lands as ready for review; only a human acceptance moves it to accepted — attempted, reviewed and accepted are three different states.
What you can use Muse for
The full breakdown
Muse is the integrator because it holds the whole picture — your memory, your files, your accounts — and can build, not just advise. The setup that gets the most out of it: keep one point of contact (you → Muse), let ClickUp be the shared handoff log, and give standing rules once instead of approving every step.
Knows you & remembers
Standing memory: your rules, people, projects, locked decisions — applied without re-briefing
Morning brief, inbox watch, calendar — the day's picture in one place
Self-check discipline: every number pulled fresh from the source before it's reported to you
Batch campaigns (like the query wave) under standing authorization, with per-item QA
Long lanes — downloads, transcription, indexing — kept moving with stall detection
Researches & verifies
Deep web research with sources, like the report behind this page
Live verification of people, companies, prices, windows — stale labels re-checked, never trusted
Best-practice first: how pros actually do it, before improvising
Builds pages & apps (Spaces)
Pages — hosted, can be published to a public link: the Master Manual, the daily brief site, this dashboard
Apps — pages with their own database that remember what you enter: trackers, planners, logs; they stay private to you
Documents too: Word, PDF, decks, spreadsheets you keep editing
Builds from your real stuff
Your Drive, Gmail, Calendar, Sheets, ClickUp, Teaching Library, workbooks
Searchable systems over your own corpus — the Asset Management System is the template
Nothing invented: dashboards show your actual agents and their actual state
The honest limits
No native phone apps; pages don't work offline
A static page can't save typed input — that's what the app kind is for; an app can't be published publicly
Live data needs a live source — this board refreshes on status events and snapshots, not magic
Deleting an artifact is permanent — Muse always checks with you first
How this gets built
Best practice, phased — from the Nov 2026 research
The dashboard is a control layer, not a chart wall. Its core object is the run record — one row per agent run, designed before the first model call — and five surfaces grow on top of it. This page you're reading is the design + live snapshot; the phases below wire it to real run data.
The run record — one row per run
Why it's there
run_id · trace_id
Connects product events, model calls, tools, retrieval and approvals
agent_name · agent_version · prompt_version
Makes regressions debuggable after a prompt edit or model change
model · input/output/cached tokens
Turns usage into cost and routing decisions
tool_calls[] · retrieval_set_id · citation_count
Separates bad agent behaviour from bad dependencies; ties quality to the exact knowledge path
approval_status · eval_scores · failure_class
Proves what was approved, rejected or timed out; makes quality searchable instead of anecdotal
The five surfaces (what grows on the run table)
Launch gates — success rate, unresolved failures, eval pass rate, approval backlog → decide whether to scale a workflow
Run explorer — searchable runs by workflow, version, model, state → debug a failed or expensive run
Trace detail — nested model / retrieval / tool / approval spans → find the exact failing step
L3 Prepare consequential actions for human approval · L4 Unattended inside narrow limits, with monitoring and stopping rules
Begin read-only; promote a workflow only after it passes real cases repeatedly
An agent trusted to draft a brief has not earned the right to touch the CRM — authority is per workflow, not per agent
Every agent gets a written charter (outcome, scope, sources, tools, evidence rules, definition of done, escalation, owner, review date) and a line in the agent register: no register entry, no credentials
WEEKS 1–2
One workflow + the run record
Pick one recurring workflow with a clear output (the query campaign is the natural first — it's already run-record-shaped)
Create the run table with the fields above; document today's baseline (cycle time, quality, human effort)
One agent, read-only or draft-only (L0–L1), with a written charter and a named business owner (Derek) and technical owner (Stephanie)
WEEKS 3–4
Coordinator + one gate
Status events wired in (started / progress / blocked / failed / completed; ready-for-review → accepted)
One consequential action behind a real approval gate — approve / edit / reject, requests expire, default is deny. This board's snapshot becomes the live view of that table
MONTH 2
Observability, evals, budgets
Langfuse (self-hosted, free) or LangSmith for traces + evals; shared trace IDs so traces survive a tool switch
A 10–20 case eval set from real tasks — golden, edge, adversarial, and approval checks (proof the agent actually stops) — run on every prompt/model change
Per-run and per-day spend caps; weekly review of cost per successful outcome, not token totals
Weekly (Derek + Stephanie): sample live work, log failures by cause, eval pass rate, rubber-stamp check (if 99% of approvals pass in seconds, the tier is wrong)
Monthly: ops review — errors, limit hits, denied approvals, cost; promote one workflow one authority level, on evidence only
Quarterly: owner review (value vs cost, keep or retire), access recertification, and test the kill switch — stop → contain → tell → fix → resume through a launch review → learn. A kill switch nobody has tested is a hope, not a control
The stack the pros converged on (Nov 2026 — verify pricing before buying)
Layer
Pick
Why / status
Agent runtime
LangGraph 1.x for anything touching money, customers or publishing; n8n as no-code glue; Claude Agent SDK if the stack goes Claude-first and file-heavy
LangGraph is the only one with checkpoint-based durable execution built in — resume from the last checkpoint after a pause instead of restarting. CrewAI is the fastest prototype but has no built-in checkpointing
Observability
Langfuse (self-hosted, MIT core; Cloud from free Hobby tier, Core $29/mo, Pro $199/mo as cited Nov 2026) — or LangSmith if you commit to LangGraph ($0 Developer / $39 per seat Plus)
Traces + prompt management + evals + dashboards in one product; emit OpenTelemetry spans with shared trace IDs
Run table + live view
Postgres / Supabase + Realtime; this dashboard as the business control layer on top
Developer tools like LangSmith Studio are IDEs, not ops dashboards — the business layer is custom, and it can start as a table + views
Avoid
AutoGen — archived (last release Sep 2025, no fixes). Helicone — in maintenance mode since its Mar 2026 acquisition; avoid for new builds
Both per the Nov 2026 framework comparisons
The honest gap: there is no published, audited case study of a founder-led courses/books/coaching business running a multi-agent fleet behind a live ops dashboard — the research looked, hard. The closest real patterns are single-agent (a creator's trained support agent answering from his own course library) or a personal AI operating system in a terminal. The documented dashboards are vendor products and engineering write-ups. The realistic move is the one above: build the run-record dashboard for this business, in phases — and become the case study. Revenue and margin figures quoted for famous one-person businesses ($15M, 94% margins and the like) are self-reported, not audited — treat them as direction, not data. Same caution for survey numbers: Gartner's “over 40% of agentic AI projects cancelled by end-2027” is an analyst projection, useful for taking governance seriously, not a measurement.