The agency that works while I sleep

This doesn't work as a chat window. It works because the AI has the same view of the business I do. Through MCP connections it reads the company email, the calendar, and the drive: every client thread, every scheduled call, every document we've traded. The task list lives in Postgres and the AI has full write access — it creates, updates, and closes tasks like any teammate. It can also reach every channel the agency operates: the website backends, Shopify through a custom app, Google Ads, the social queue, the scraping and SEO tools.

The AI also has a name and a seat on the team. Cassi Claw is the agency's AI operations agent, with her own account and her own email address. Every message she sends passes through a wrapper that checks the recipient against an allowlist and writes an audit line. Clients know Cassi is an AI. That's deliberate: a transparent team member, not a ghostwriter behind my signature.

That's the whole trick. An AI that knows the schedule, the clients, and the task list, and can act on all three. Everything below follows from it.

Here's one rotation from a real week: the Friday hand-off, the weekend shift, and the Monday that follows. Tuesday runs the same shape, and so does every day after it.

A few beats worth slowing down on. The call agendas arrive built: the last seven days of cost, impressions, clicks, and conversions by channel, with a first-pass analysis. I add my own notes on top — a seasoned operator's read, a little less textbook than the model's — and that layering is deliberate. On the call itself, the AI is a live analyst: when a client asked about a week-over-week conversion dip, a quick statistical check showed the dip was noise, and gave us the exact number that would mean it wasn't. And afternoon work compounds, because every ad-copy rewrite is logged into the optimization loop as a hypothesis and graded against real outcomes at day 3, 7, and 14. The account's history reads like a lab notebook.

The overnight engine is the orchestrator documented here: draft PRs only, seven safety layers, budget caps, and a journal entry for every task with its cost. The blogs come from the same discipline as the skill libraries: forced research, self-grading, brand red-lines read verbatim before a word is written.

The highest-value use of the loop is a campaign launch. The marketing calendar is under continuous analysis, so the system already knows a three-day free-shipping sale is due next week — best channel mix, best timing, best offer, all from the data. From that single calendar entry, the night shift builds the whole launch. The email campaign is drafted against a visual-richness rubric with the brand's red-lines enforced. The social graphics go through a render-and-critique loop where the first version never ships. And the storefront pieces — a banner, a campaign landing page, a new section — are composed on a staging theme from a vetted section library.

Even the go-live is scheduled. My campaign engine deploys timed content changes through Shopify's metaobjects and records rollback data for every operation, so the sale turns on at midnight and turns itself off three days later without anyone touching the theme. A human approves the whole package in the morning — nothing publishes without that. But the building happened while everyone slept. For a store running a promotional calendar, that's a launch costing a review instead of a team-week.

The tasks above are examples. The system is the library: 10 playbooks the overnight engine can run (dev changes, cross-repo ports, blog posts, emails, data fixes, research briefs, recurring reports, ads reviews) and 8 client-facing skills (campaigns, graphics, social calendars, landing pages). Each one opens by declaring exactly when it fits and carries its own guardrails. One playbook is simply "stop and ask" — the designed escape hatch for anything ambiguous. A shared task format makes all of it executable without me in the room.

That's what makes it plug and play. When a new kind of work shows up twice, it gets codified: when it fits, what done looks like, how it self-checks, what it must never touch. Once vetted, it joins the library, and the night shift can do it with our eyes closed. The library improves itself, too — a weekly pass has the AI read its own journal and propose edits to the playbooks based on what actually happened. The list of things this agency does unattended gets longer every month, and the loop never has to change to absorb it.

The AI is staff, and it's managed like staff. Everything outbound is draft-first: emails queue for approval, pages land unpublished on staging themes, PRs arrive as drafts that cannot merge themselves. Cassi's sends are allowlisted per recipient and logged per message. The overnight engine runs behind budget caps, a kill switch, stuck detection, and a verification gate that re-runs the tests and locks any PR whose tests didn't really pass. Every content skill reads the brand's red-lines before writing and self-critiques before handing off.

And it all leaves a paper trail: a journal entry per overnight task with approach, outcome, and cost; an audit line per email; a hypothesis record per ad change. When I open my laptop, I can reconstruct exactly what my night shift did and what it spent doing it — which, at three or four dollars per completed task, is the least dramatic number in this report.

This is a boutique operation: five active clients, one human, and a system tuned to that scale — though the architecture is the same one I'd run at ten. The launch stack composes capabilities that are each real and separately documented, and by design a human approval sits in the middle: nothing on this page publishes, sends, or spends on its own. The costs are measured, not estimated — they come from the overnight journal, which records what every task actually cost.

What it can say: an agency's operating layer — the planning, the prep, the content, the follow-through, the bookkeeping of who promised what — can be run with AI as genuine staff today, on a library of vetted moves that only grows. The human's job concentrates into the two things that deserve a human: judgment and relationships. Everything else gets a night shift.