THE AGENTIC C-SUITE: HOW TO RUN A FULL EXECUTIVE TEAM IN CLAUDE CODE

A six-person business cannot afford a CFO. Or a CMO, a CTO, or a chief of staff. So the work those people would do, watching the numbers, catching the drift, keeping the machine pointed at the goal, mostly does not get done. It gets done in a panic, once a quarter, when something has already broken.

We stopped accepting that. We now run a full executive team, and it costs tokens instead of salaries. Each executive is a Claude Code session with its own memory, its own Slack channel, and its own loops. This is written from the terminal we built it in, not from a whitepaper.

From "wear the hat" to "run the session"

Six months ago the clever move was to ask the AI to wear a hat. Be my CFO for five minutes. Be my CMO. It worked, and it beat nothing, but it forgot everything the moment you closed the chat.

The version that actually runs a business is different. You do not ask for a hat. You open a dedicated session per executive and leave it running. The CMO session knows it is the CMO. It knows what converted last week, which loops it owns, and where the gate is. It reads its channel every morning and tells you the three things worth your attention. You reply, you approve, and it gets on with it.

Boris Cherny, who built Claude Code, put it plainly: his job now is to write loops, not prompts. The agentic C-suite is that idea aimed at a business instead of a codebase.

The shape of it: four sessions, one brain

We run four sessions. CEO, CMO, CRO, CTO. Each has one job.

• The CEO session is chief of staff. It reads the whole business at altitude and hands over the three priorities only a human can action today. One page, one coffee.
• The CMO session owns demand: content, social, search, and the market's own words.
• The CRO session owns revenue: the pipeline from a cold connection to a booked meeting.
• The CTO session owns the wiring: the platform, the connectors, the security, the build.

They do not argue, because they read and write the same brain. One database that is the single source of truth, plus a set of memory files that get sharper every week. A competitor with the same AI starts month one with an empty brain. Ours knows more about how we work in month twelve than it did in month one. That accumulation is the moat.

The three loops that make a session work

A session is not a chatbot. It runs on three loops, and the third is the one most people miss.

The daily loop is the morning kick-off. It pulls the brain, reads its channel, and surfaces what needs a decision. For the CEO that is three priorities. For the CRO it is the approval queue and what moved in the pipeline overnight.

The hourly loop keeps the session live through the day. A new reply lands, a draft is ready, a connector goes down: it surfaces the thing, and otherwise stays quiet. Silence is a feature. You do not want four sessions chattering.

The delivery loop is the engine. This is the one that finds work, does the work, checks its own work, and delivers it, then goes and finds the next thing. Not "here is a report", but "here is the thing, built". The CTO delivery loop chews through the build backlog. The CMO delivery loop drafts the week's content. It stops at exactly one place, and that place is the whole game.

The one rule that makes it safe to run

Automate everything up to the point of judgement. Stop dead at anything a customer will see.

Everything that only touches data runs unwatched: the research, the enrichment, the reviews, the health checks, the morning brief. Everything that touches a person waits for a human yes: every message, every post, every published page, every deploy. When a coding loop writes bad code you revert the commit and nothing left the building. When a sales loop writes a bad message it lands in a real inbox, in your name, and you cannot revert a first impression. So the delivery loop drives itself right up to the send button, and then it stops.

Get that rule right and the maths changes. You are not hiring four executives. You are approving the work of four executives, which takes a coffee, not a payroll.

What it actually looks like running

We run this on our own business first. The pipeline session drafts and stages every message and reliably keeps around eight meetings a week moving, without a single unapproved message ever leaving the building. The rejection rate on drafts starts at about one in ten and falls to near zero within a week, because every edit at the gate teaches the swarm. The build session works a backlog we used to work by hand.

Here is the honest part. The constraint is no longer the AI. It is us. On most mornings there are more approved-in-principle drafts waiting than one person can clear before lunch. That is a good problem. It is the problem of a business that suddenly has a full executive team and a founder who is still the only one allowed to say yes.

How to start: one session, not four

Do not open four sessions on Monday. That is how you get a confident conveyor belt for bad work, four times over.

Start with the CEO session. It is data-only, zero customer risk, and immediately useful: a one-page brief every morning that tells you what to look at. Write down what a good brief looks like, so the session has a standard to hit. Add a verifier before you add autonomy. Then, and only then, turn on a single delivery loop and point it at something small and checkable. Widen to the next session when the last one has earned it.

The firms that get burned start at the delivery loop. The firms that compound spend a fortnight on the brief.

The bottom line

You cannot afford a C-suite. You can afford the tokens. Your morning starts with a virtual executive team handing you a short list, a set of loops covers the endless surface underneath, a verifier keeps them honest, and you stand on every exit a customer can see. You decide. The swarm delivers.

Want this built and run on your own accounts, with you approving every move? That is what we do at Neon Gorilla. AI Sales, Search and Social, on your accounts, live within seven days.

Sources

Boris Cherny is the creator and head of Claude Code at Anthropic. His remarks on loops are from his 2026 public talks and interviews, collected at howborisusesclaudecode.com and written up in The New Stack's "Loop Engineering". Addy Osmani's warning that a loop can become "a very confident conveyor belt for bad work" is from the same piece.