Running 10 Social Clients Solo: The AI Agent on a Leash

I run social media for 10 clients by myself, using an AI agent that is never allowed to publish anything. That sentence has two halves and most people only hear the first one. The agent drafts. I approve. The queue does the rest. Take away the approval step and this whole setup falls apart in a week, probably during a client’s product launch.

Ten clients solo is already above the norm, before an agent enters the picture

A Databox survey of social media agencies found that nearly 70% keep account managers under 10 clients, with most clustering between 5 and 8. Karl Sakas, who has advised hundreds of agencies, puts a healthy range at 6 to 8 easy clients or 2 to 3 needy ones. I sit at 10, alone, and I did not get there by working faster. I got there by removing myself from one specific step: the first draft of every routine post.

The agent I use connects to PostSider over MCP and does exactly one class of work well: producing a first draft per client, per channel, based on what that client’s brand voice actually sounds like. It does not decide what ships. It does not touch a client’s live feed. It writes, and then it stops.

The agent drafts per client, into a status that cannot publish itself

Every post the agent creates goes in as draft, one of three statuses in PostSider alongside schedule and now. Draft means exactly what it sounds like: the post exists, nothing about it is live, and nobody outside my own dashboard has seen it. I run 10 separate organization workspaces, one per client, each with its own API key, so the agent working on Client A’s drafts has no path to Client B’s channels even if a prompt went sideways.

That separation matters more than people expect. An agent juggling 10 brand voices in one context window is exactly the setup where a stale detail bleeds from one client into another. Ten keys, ten workspaces, zero shared context, is the boring fix.

Nothing publishes without the morning approval ritual

Every weekday morning I open the queue and go client by client. Fifteen to twenty minutes, coffee still warm, reading what the agent wrote overnight against what I know about that account. Approve, edit, or reject. Nothing more complicated than that, and nothing less disciplined either.

This is the step people skip when they get excited about agent autonomy, and it is the one step I will not skip for a client account. My own accounts, sure, I experiment more. A client’s brand voice going out wrong is not my mistake to absorb quietly. It is a call I lose.

Once approved, the queue and Smart Slots do the placement

Approval is a decision, not a scheduling problem. Once I approve a draft, PostSider’s day-aware posting queue and Smart Slots pick the actual publish time per channel, so I am not manually hunting for a free slot across 10 clients and however many connected networks between them. That is the part of the job that used to eat an entire afternoon and now takes the same fifteen minutes as everything else, because the placement logic runs on its own once a human has said yes.

What never touches the agent, on any client, ever

Three things stay fully manual, no exceptions: strategy, replies to comments or DMs, and anything that smells like a crisis. Strategy is a conversation with the client, not a drafting task. Replies are real-time and personal in a way a batch-drafted post is not. A crisis needs someone who can read the room in the next five minutes, not the next morning’s queue.

I say this plainly because the pitch for agent-run agencies always undersells this part. The agent is good at volume. It is useless at judgment. Confusing the two is how an agency ends up with a competent-sounding apology post that still says the wrong thing.

The honest number on time saved is smaller than the hype

I do not have a rigorous before-and-after study of my own 10 clients, so I am not going to hand you a made-up number dressed as personal data. What I can point to is external research: automation tools save an average of 6 hours per week per person, according to ImpactPlus, which adds up to more than 300 hours a year. That is the drafting time back, not the judgment time, and it is consistent with what running this setup actually feels like. The agent buys me back the hour I used to spend staring at a blank composer for each client. It does not buy me back the fifteen minutes of judgment every morning, and it should not.

Start at one client, then add the leash before you add the ninth

If you are trying to build something like this, do not start with 10 clients and an agent on day one. Start with one, run it at draft-only for a couple of weeks, and watch what the agent actually produces before you trust it with a second brand voice. I wrote more about that gradual trust curve in how to let an agent post without waking up to a disaster, and about the broader shape of agent-run marketing in where AI agents fit into social media strategy. The mechanics scale the same way from one client to ten. What does not scale is skipping the approval step to save fifteen minutes. That is the fifteen minutes that is doing all the work.