7 min read ·
AI Marketing Agent: What It Should Do Alone
Quick answer
What an AI marketing agent should do on its own, and what it should leave to you: a permissions list for makers, split into what to allow and what to keep.
An AI marketing agent is a program that works one marketing channel on a schedule. It reads what is happening there, picks what is worth acting on, and writes what a person would have written. That definition leaves the hard question open. Once a program can read your analytics and hold your X account, which of those steps should it take alone?
We build Kwern, which runs 14 of these agents for indie makers, so our answer is written into how the agents run. This is the permissions list we work by, with the reason behind each line. Use it to judge any agent, ours included.
Three rules the rest follows from
Our agents run by three design rules, and most of the permissions below fall out of them.
One agent per channel. The Reddit agent works Reddit. The Apple Ads agent works the Apple Ads account. Each runs only the marketing skills its template allows, so the Reddit agent cannot decide to rewrite your App Store subtitle and the ads agent cannot decide to write a blog post. A narrow agent is easier to trust because you can predict what it will hand you.
Approval on every action. Every post, reply, listing change, pull request and bid arrives as a proposal, and nothing reaches a channel or an ad account until the owner approves it. We argued the case for that in AI CMO for indie makers. This piece is about everything the agent may do before that point.
Your own model key. The model runs on the maker's provider key at the provider's price, and Kwern records what each run cost. That rule matters more for permissions than it first looks, and we come back to it below.
What an AI marketing agent should do on its own
Wake up without being asked
An agent you have to open is a chat window with a job title. It gets three good weeks and then becomes a tab you stop opening. An agent should run on a cadence set by how fast its channel moves: in Kwern, Reddit and X replies run daily, X posts every three days, the SEO crawl weekly and the GEO probe every two weeks. Waking up is safe to automate because it changes nothing outside your account.
Read everything it has been given
Reading is the slow part of marketing and the part worth handing over: the week's reviews, the search terms behind ad spend, the threads where someone asks for what you built, the pages Google crawled. An agent should read through the connectors you gave it, and the public threads it can reach, with no prompt from you.
It should also say when it cannot read. An agent missing the connector it needs should stop and name what is missing. A run that carries on without its data writes from guesses, and a guess in a draft looks exactly like a fact.
Decide what to leave alone
A good share of an agent's judgement goes into what it does not draft. Our reply agents answer only posts and threads under 24 hours old, because a reply to a stale thread reads as spam to everyone who finds it later. The Reddit agent stops drafting while an earlier draft still waits for you: a backlog means you have not had time to read, and more drafts make that worse. A run that concludes nothing is worth doing today has done its job.
Draft the finished thing, with its reason
A proposal should be ready to publish as it stands: the exact text, the exact bid, the exact listing value, the hour it would go out. Beside it goes one sentence saying why, naming the thread it answers, the keyword it targets or the search terms behind a bid. Drafting is safe to automate because a draft costs nothing until someone acts on it. The reason is what lets you decide in two minutes rather than redo the research.
Learn from what you did with the last draft
When you edit or reject a proposal, the next run should know. Kwern keeps a product memory, distilled from approvals and results, and every agent reads it before it writes. A rule such as "open X posts on a plain fact about what we did" came from edits the owner made to drafts, and the agents follow it now without being told again.
What it should never do on its own
Publish or spend
Every other line on this list protects this one. An agent that publishes has to be right every time. One that proposes only has to be right often enough that reading its drafts costs less than writing your own. In Kwern there is no autopilot setting, and no mode where an agent spends first and reports afterwards.
Open something new
The Apple Ads agent and the Google Ads agent propose bids, budgets, keywords, negatives and pauses. They cannot create a campaign or an ad group: a new campaign starts in a chat the maker opens. Their instructions keep each move within a fifth of the current value unless the proposal says why, and every run reads the monthly budget you set. So a mistake stays small enough for the next run to undo. An agent allowed to create campaigns can make a large mistake in one step.
The same logic decides which agents run at all. Of our 14, eight are on the day you sign up. The two ad agents start off because they spend your money, the Influencers agent because it writes to people, and the Writer because an article nobody publishes is a day of the model's time for nothing.
Post where an automated account has no standing
On Reddit and Hacker News our agents never post. The Reddit agent drafts, and you post from your own account; replies on X work the same way. Both communities treat an automated account as one, and when an account is banned, the maker is the one who loses it. Where a service does not let Kwern write, approving is not even a button: you get the text and the link.
Pretend to be someone else
An agent should post as you, from your connected account, or stay silent. It should never run a second account, and never write to a person you have not been shown first. Our Influencers agent finds up to five people a week whose readers are your buyers and drafts the message to each. You send it.
Make a number up
Every draft should be written against two sources: the Foundation doc, a written description of your product, audience and voice, and the data your connectors returned. Where only the maker knows a detail, the Influencers agent leaves it in brackets for you to fill in. An agent that fills gaps with plausible figures will one day put a false claim under your name, and you will be the one who answers for it.
Step outside its channel
This is the first rule seen from the other side. An SEO agent that notices your App Store listing could be better should leave it to the agent that reads the listing with the listing's data. When every agent may touch everything, two of them will sooner or later propose conflicting changes to the same field.
Why the model key belongs on a permissions list
Most AI marketing tools bill the model as credits. A credit balance hands one permission to the vendor: how much reading and drafting your marketing gets this month is set by a number you cannot convert into tokens. When it runs out mid-month, the agents stop, whatever the channel needs that week.
On your own key, the model is a line on your provider bill at the provider's price. Across the 11 models Kwern vets, input runs from $0.10 to $10 per million tokens, and which end you sit at is your call. Kwern records the cost of every run, so you can see which agent is expensive and switch it off. The agent may spend tokens on reading and drafting; you see exactly how many, and you set the ceiling by choosing the model.
Eight questions to put to any AI marketing agent
Ask these before you connect an account, ours included:
- Which of its actions can reach the public or spend money without my click, and could anyone switch that on in a setting?
- What does it do when a connector it needs is missing?
- Which channels will it refuse to post to, and does it say why?
- Can it create campaigns, accounts or messages to strangers, or only change what already exists?
- How far can a single change move a bid or a budget?
- Does each proposal carry its reason, with numbers I can check?
- Who pays for the model, at what price, and can I see the cost of each run?
- Can I read the instructions it runs before it runs them?
On the last question, the 76 skills our agents run are open files on the skills page, readable without logging in. The agents page lists all 14 with their cadence and the limits above, and the pricing page has the three plans, from $29 a month for one product, with the model on your own key.