Fleet SOW — the agent system

The canonical map of who does what — three functions, who is human and who is AI, five roles, one scoreboard, and every daily and weekly duty.

Updated 2026-08-01

1 The three functions

The whole business is three things. Nothing else counts as a function.

Function	Leading — what you control	Lagging — the result	Who does the human part
Marketing	Content shipped	Leads	Akash records the content
Sales	Conversations started	Appointments booked	Akash holds the conversations and closes
Delivery	Tickets resolved, and how fast they get resolved	Cash flow	Akash makes the judgment calls on tickets

Only three things need a human at all:

1Make more content.
2Start more conversations.
3Resolve tickets faster.

Everything else runs in the background, automatically — onboarding, campaigns, media buying, the mechanics of running ads, the mechanics of outbound, and reporting. If a job is not one of the three above, it should not land on Akash's plate.

Push on the leading number today; the lagging number tells you a week later whether the pushing worked. Never plan against a lagging number — you cannot move it directly.

2 Human roles vs AI roles

Akash does	AI does
Records the content	Writes the scripts, edits, schedules, posts it, reports on it
Holds the conversations and closes	Starts them, replies, qualifies, chases, books the time
Makes the judgment calls on tickets	Opens, sorts, drafts the answer, chases the client, closes the loop
Says yes or no at the three gates	Prepares everything right up to the gate, then stops

The three gates — the only places work is allowed to wait on Akash:

1Spend — money going out, or an ad going live.
2Publish — anything going public under the brand.
3Outbound sends — any message going to a real human as Akash.

If something is waiting on Akash and it is not one of the three gates or one of the three human jobs, that is a bug in the system, not a task for him. Fix the automation instead of adding to his list.

3 The one-sentence model

Skills are the how, scheduled jobs are the when, the worker is the hands for one-off work, and the named agents are accountability bundles on top — each owns one number, none is a separate always-on program.

You steer from exactly two places: Telegram (talk, read reports, approve) and Linear (queue work).
Everything an agent produces lands as a message to you or a draft for your QA. Nothing sends externally, spends, or publishes without your go.
Every result is recorded in one shared memory (the brain database). Telegram is a pane of glass, never the record.

4 Your two surfaces

Surface	What it's for	How you use it
Telegram — St. Pierre HQ	Reports find you; you reply to act	Topics: Daily Reports · Alerts · Notes · Approvals · Agent Work
Linear board	You queue and promote work	Add the `agent-ready` label to a card = the worker may claim it

The old iMessage alert path is dead — it delivered to the Mac, not your phone. Telegram HQ replaced it on Aug 1. iMessage remains only as a last-resort trace on the Mac.

5 The scoreboard — the loop engine

The loop engine never does any work. It sets the bar and keeps score, per department, every week.

1Sunday evening: drafts next week's numeric targets for each department. You review; only your explicit word locks them.
2Friday afternoon: compares target vs. actual, writes down learnings, proposes next week's bar (aims to tighten toward a 75% hit rate).
3Mid-week: any number going off-track becomes a line in your morning brief.

Rite	When	State
Sunday targets draft	Sun 6:02 pm	Running — lands in Daily Reports
Friday reflection	Fri 3:10 pm	Paused — awaiting your word since Jul 31
Off-track routing	mid-week	Runs with the morning brief

The difference from the fleet: agents produce outcomes, the loop engine judges whether outcomes hit the number. Workforce vs. scoreboard.

6 The five roles and their numbers

Each role is a bundle of scheduled jobs, skills, and cards — grouped under a name so one thing owns one number. Every role sits under one of the three functions; nothing lives outside them.

Marketing: Morgan, plus the whole content machine behind him.
Sales: Alex, plus every piece of outbound.
Delivery: Jordan and Casey.
Background supervision: Sam. Not a fourth function — the thing that watches the other three and shouts when one stops.

Function	Role	Owns this number	State (Aug 1)
Background supervision	Sam — supervisor	Nothing stalls silently	Half-live: morning brief running, worker paused
Sales	Alex — B2B sales + outbound	Qualified launch calls booked	Staged bot; booking wire missing
Delivery	Jordan — B2C delivery	Qualified homeowner handoffs	Live per client; needs templating
Delivery	Casey — client care	Job outcomes captured (the revenue loop)	Not built — the #1 build
Marketing	Morgan — marketing + content	Daily pack shipped + cost per lead in range	Pieces exist; no single owner

Sam (background supervision) — the one voice you talk to in HQ. Daily: composes the 7:30 morning brief from live data (pipeline, open alerts, stuck work). Continuously: the watchdog chain escalates anything that goes quiet. When the worker resumes, Sam's reports cover what it did. Sam owns no function number — he exists so the other three never fail silently.

Alex (sales) — answers inbound contractors, qualifies, books the launch call, and runs the outbound mechanics underneath. Drafts only until 50 clean drafts are reviewed. Escalates price talk, anger, or unclear fit to you. Serves the leading number: conversations started. One build step remains — wiring the calendar booking action.

Jordan (delivery) — the homeowner-facing qualifier we deliver to every client. Level King's is live and is the golden copy. Next: freeze that config as a per-niche template so a new client's bot is a clone, not a rebuild; add a nightly check for homeowner frustration, not just bot uptime.

Casey (delivery) — owns everything after a lead is handed to the client: confirming delivery, chasing the client for outcomes with two-question texts, the daily report to you and to the client, review requests, reactivation. This is the role that turns resolved tickets into cash flow. Does not exist yet. Gated on ~30 minutes of your clicks (pay the small ad balance, one real test message, four GHL settings).

Morgan (marketing) — one daily rhythm over the existing creative machinery: build the day's ad pack as paused drafts, post a five-line spend note, recommend kills and scales, weekly competitor sweep. Serves the leading number: content shipped. All Meta activation stays behind your approval gate, always — that is the spend gate.

7 What runs daily

Time (ET)	What happens	Role	Where it lands
6:20 am	Yesterday's sessions archived to the brain	rails	silent
7:30 am	Morning brief composed from live data	Sam	Daily Reports
7:45 am	Ad pack drafted, paused, never activated	Morgan	local drafts for your OK
every 5 min	Board changes pushed to Linear	rails	silent
every 10 min	Health checks: is anything broken or stalled?	Sam (watchdogs)	Alerts, only when something breaks
9:00 am	Independent cloud cross-check of all rails	Sam (watchdogs)	Alerts
6:00 pm	Daily-three discipline score	scoreboard	Daily Reports
10:10 pm	Check that every automation is registered and watched	Sam (watchdogs)	silent

8 What runs weekly

When	What happens	Where it lands
Friday 3:10 pm	Loop engine reflection: target vs. actual, learnings, next bar (paused)	Daily Reports when resumed
Friday 4:00 pm	Weekly grade	Daily Reports
Sunday 6:02 pm	Loop engine drafts next week's targets	Daily Reports
Sunday shutdown	Week review + next week's plan through the shutdown ritual	your shutdown report

9 The worker — scope rules

The worker is one generic laborer for one-off judgment work. It has no identity and owns no number.

Claims only Linear cards carrying the agent-ready label (core lane) — enforced in code, verified Aug 1
At most 3 cards per half-hour cycle, with a hard spend ceiling per run
Drafts and builds only; parks everything at "waiting" for your QA
Reports every result to the Agent Work topic in HQ
Never activates ads, never messages a human, never deploys, never touches money or credentials
Never marks work done on your behalf
Never claims anything outside Linear — the old side-channels are closed (verified zero remain)

The worker is paused. Resume is one command, gated on your word. On resume it would claim the oldest three queued cards first — pull the heavy "ship ads batch" card out before resuming, or expect it to burn most of a cycle there.

10 The three primitives

Every role above is assembled from the same three parts. Nothing else should exist.

1Scheduled jobs — the when. Cron on Hermes for reports and rites, n8n for event-driven syncs and webhooks, launchd for local renders and watchdogs. Mostly no AI; where judgment is needed, a small AI step runs inside the job.
2Skills — the how. The recipe library: ad building, drafting, video production, audits, onboarding. Any layer can invoke a skill; skills change only through the skill-creator process.
3The worker — the hands. One-off work you promote via Linear. Everything recurring belongs to a scheduled job instead.

The July lesson, now policy: no long-running AI programs beyond the gateway. Hermes was silently dead July 3–28 and nothing noticed. Watchdog before workload, always — its death now alerts your phone within the hour.

11 Build states and gates

Item	State	Gate to change it
Telegram HQ + all report routing	Live (Aug 1)	—
Hermes on the company's own AI key + watchdog	Live (Aug 1)	—
Morning brief, daily score, weekly grade, Sunday targets	Live (Aug 1)	—
The worker	Paused, scope locked to Linear	Your word: "resume the worker"
Friday reflection	Paused	Your word
Alex's booking wire	Unbuilt	None — agent work, ready to run
Jordan templating + frustration check	Unbuilt	None — agent work, ready to run
Morgan consolidation	Unbuilt	None — agent work, ready to run
Casey (the #1 build)	Unbuilt	~30 min of your clicks first
Organic publisher	Parked	Your word — and its switch is set to LIVE publishing; flip to dry-run before any resume
Newsletter draft	Parked	Your word
Slack retirement (formal)	Waiting	After a week of Telegram reports proving out (~Aug 15)

12 For agents — where things live

Reference pointers for whoever works on this next. Safe to skip if you're not an agent.

Worker code: 01-Projects/St_Pierre_Digital/agent-board-rail/hermes_runner.py (claim gate: Linear-linked + agent-ready + core lane)
Escalator: 01-Projects/CommandCenter/watchdog/escalate.py → HQ Alerts topic; Hermes health: hermes_heartbeat.py, launchd com.stpierre.hermes-heartbeat
Telegram send path: brain.telegram_hq_send(topic, text); topic ids live in the vault row (telegram / progress_tracker), key hq_topics
Hermes cron jobs: ~/.hermes/cron/jobs.json; report jobs deliver to telegram:-1004401888570:7
Loop engine: 01-Projects/StPierre-GTM/loop-engine/loop_engine.py (project Sundays, reflect Fridays — reflect paused)
Fleet contracts in full: 01-Projects/St_Pierre_Digital/SOWs/sp-agent-fleet-enriched-plan-2026-07-28.md (supersedes all earlier Viktor-era docs)
Cutover plan: _scratch/deliverables/2026-08-01/sow-cutover/index.html (the 14-item off-the-terminal SOW)
Never: getUpdates polling, setWebhook, claude -p rebuilds, pip inside the Hermes venv (use uv)