The Agent Manager Playbook

Part 1: Carson's Verified System

Founder of Treehouse (taught ~1,000,000 people to code), now running Untangle (AI divorce agent for family-law firms) as its only employee. Revenue 4x'd the month of recording.

22 to 25

Pull requests per day, spiking to 40

10 to 20

High-stakes decisions before lunch (was 2 to 3 per day pre-agents)

~50%

Of his work done from his phone

$20,000

Token bill last month ("not viable"), targeting $5,000 per agent-employee

The three automations he actually runs

End-to-end signup test. A real browser does a real signup, 3 times a week. Costs about $60 in tokens per run. Breakage is found by the robot, not by a customer.
Production watchdog, daily at 9:00 AM. Parses the day's activity into one summary with links straight into his admin UI. "I can't keep track of it mentally by just looking through logs." No dashboard to open, the digest comes to him.
Grace, the self-improvement loop. Grace is his paralegal agent that talks to attorneys and clients. Every day an automation pulls yesterday's conversations, grades each one against a rubric he wrote, and anything below threshold kicks off a child session to fix it. "It is shocking to me the number of small, fine UX details that have been picked up by this loop and have been fixed."

His generalizable recipe, stated explicitly

Pick the part of your product you want to self-improve. Write down how an agent judges whether a given interaction was good or bad. Run it daily. Auto-fix anything below the bar.

The management patterns

Parent/child model split. "Use a Fable thread as the manager and ask it to spin up children... it's not viable to run a super premium model all the time, but have a very intelligent parent run children" on cheaper models. Same as our 10-80-10 split.
The 25-minute cadence. He checks agents and makes decisions in 25-minute sweeps, not continuously. The point is mental rest: batch the decisions, don't hover.
The manual key gate. Agents never hold production write keys. They must ask; he pastes the key in for that one task. "Production writing is dangerous, and your agents will do something bad."
Don't build your own software factory. "If you're trying to build a software factory for yourself, stop." One-person shops should ride the $200/month harnesses and let the agent labs do model routing.
The honest cost. He does not sell this as freedom: you "work a lot more, not a lot less," and the whole picture "is going to change every 3 to 6 months."

Part 2: How This Fuses With Loop Engineering

Carson never says the word "loop engineering," but all three of his automations are textbook loops. Map them onto the six-part anatomy from the Loop Engineering Playbook and you see he's running the exact same machine, pointed at the BUSINESS instead of the codebase.

The Grace loop, in loop-anatomy terms:
TRIGGER -> daily
DOER    -> pull yesterday's real customer conversations
CHECKER -> grade each against a written rubric (the human wrote the rubric,
           the agent applies it: worker isn't grading its own homework)
STOP    -> all conversations graded; fixes spawned only below threshold
MEMORY  -> the rubric + the grade history
SKILLS  -> child fix-sessions on cheaper models (parent/child = 10-80-10)

The insight that matters for us: the loop playbook taught WHERE loops come from (repeat asks) and HOW to build them safely (checker, caps). Carson adds WHAT to point them at: real customer interactions, graded nightly. Code loops make you ship faster. Interaction-grading loops make the product better while you sleep, and they find work nobody would ever assign: "there are things that I just wouldn't bother doing."

Where we're already ahead of Carson (verified in the analysis)

His production-key gate is manual copy-paste discipline; ours is mechanical (PreToolUse hooks that block whole classes of mistakes). He has no described memory or corrections system; ours compounds every mistake into a permanent rule. His agents do engineering only; ours run SEO, outbound, reputation, support, and content across five brands. And his Slack approvals are our XaviDash + iMessage bridge, which already exists.

The honest strategic read

Carson's 50x is downstream of product-market fit: real customers generated real demand, and shipping capacity was his bottleneck. Our bottleneck is revenue, not shipping speed. So we import his loops where they touch revenue (call quality, checkout reliability, lead handling), not to inflate a PR count nobody is paying for.

Part 3: What We Adopt (Mapped To Our Stack)

Five imports, each one wired to assets we already own.

Carson's pattern	Our version	Ingredients we already have
The Grace loop	Nightly loop pulls the day's CallSetter voice-agent calls and demo interactions, grades each against a rubric built from the script checklist, queues fixes for anything below the bar. Voice-agent edits stay surgical and approval-gated per our laws. Same recipe later for RRC bookings and GH nurture replies.	Call recordings, script-checklist.ts, the voice-agent SOP, sales-manager-analyzer skill
9:00 AM watchdog digest	One morning digest that consolidates ALL our monitors (GH funnel, deploy ledger, A2P status, review machine, bounce alerts) into a single message with direct links. Today those are scattered outputs nobody reads until something breaks.	All the monitor crons, the brain dash, iMessage bridge
E2E flow test	Scheduled real-browser runs of the money paths: GH checkout (card + crypto), Mold Scanner scan purchase, RRC booking, CallSetter demo booking. Screenshot proof, text on failure.	agent-browser, Playwright, the checkout tunnel knowledge
Parent/child split	Formalize in every loop: smart model plans and reviews, cheap models grind. Already our subagent routing law, now a hard default in loop templates.	subagent-model-routing rule
Phone-first decisions	Carson does half his work from his phone. Our decision queue routes to the iMessage bridge: morning digest + approval asks land as texts, you clear 10 decisions from anywhere. The 25-minute batching rhythm instead of terminal-hovering.	iMessage bridge (live), XaviDash approvals write-path

Part 4: What We Skip, And One Tension To Hold

Not importing

The 22 to 40 PRs/day target. Merged-PR volume is a vanity metric for a pre-revenue portfolio.
Cloud VM migration as a 50x play. Worth doing for reliability only (our Mini has real memory-pressure scars), not as an output fantasy.
A custom model-routing engine. Carson's own words: at one-person scale, stop. The agent labs do this for us.

The tension to hold consciously

Carson's strongest opinion: "I don't understand why anybody would build their engineering on a frontier lab stack... they are not incentivized to make it reasonable for you long term." Our standing rule mandates the subscription and forbids metered keys, which is exactly the posture he calls a trap. Both are right: our rule controls cash burn today, his warning is about pricing power tomorrow. No action now, but if subscription pricing turns, the independent agent labs (Cognition, Cursor, Factory, Amp) are the documented escape hatch.

Part 5: The Build List (In Order)

Merged with the loop playbook's upgrade backlog, ordered by revenue impact.

Independent judge pass on loop completions. From the loop playbook. The safety layer everything else depends on.
CallSetter Grace loop. The single highest-leverage import: nightly call grading against the script rubric, fix queue for approval. Revenue-adjacent (better calls = more booked demos).
The 9:00 AM digest. All monitors, one text, direct links. Kills the scattered-logs problem the same week.
Money-path E2E loops. GH checkout, Mold Scanner purchase, RRC booking, CSAI demo booking, each on a schedule with screenshot proof.
Phone decision queue. Approvals and digests through the iMessage bridge so decisions batch on your 25-minute rhythm from anywhere.