Blog › Product
Autonomy needs a leash — here’s ours
Guardrails, recorded stances and human escalation: how agents stay on-brand without you watching every draft.
· 4 min read
Full autonomy with no guardrails is a liability, not a feature. Real autonomy means agents act on their own inside boundaries you trust.
We build software that can post under your name, reply to your customers and prepare job applications. So the question we get most is the right one: what stops it doing something you would not have done? This is the answer, in detail.
Three layers of control
Brand-voice grounding keeps tone on point. Recorded stances keep your positions consistent across channels. And human escalation routes anything regulated or sensitive to a person before it ships.
Trust is earned one on-brand post at a time — never assumed.
Layer 1: grounding
Every draft starts from your brand voice: a tone, an audience, do and don't lists, and samples of your own writing. An agent that has read how you write produces drafts that need less correcting, which matters, because a draft that is nearly right gets a proper read and a draft that is far off gets rubber-stamped or abandoned.
Layer 2: stances
You record what you believe: on pricing, on competitors, on anything you want handled the same way every time. Each stance has a topic, your position, and optionally the evidence behind it. They are given to agents and to the reply drafter as guidance to respect.
Each stance also has a risk level. Mark one High risk and it changes from a preference into a tripwire.
Layer 3: escalation
When a high-risk topic comes up in an inbox reply, the draft is flagged for a person instead of going out. And when the reply drafter decides it cannot answer with confidence, it escalates: the item is marked, moved to the top with a red border, and given a short, safe holding reply instead of a guess.
A system that says "I am not sure, over to you" is worth more than one that is always fluent.
The default is held
The three layers reduce bad drafts. The default handles the ones that remain.
Every agent starts in Draft first. Its posts wait in the Publish Queue marked awaiting approval. In that state a post cannot publish itself.
Two details matter:
- Changes go back for approval. Edit the text, the picture or the video on an approved post and it returns to review.
- Asking is not approving. A run you request through the in-app assistant is held for approval even if that agent normally publishes on its own.
You choose per agent
You choose per agent: approve everything, or let it run. The leash is yours to lengthen.
There are three settings on every agent's card:
| Mode | What happens |
|---|---|
| Draft first | Drafts wait for your approval. The default. |
| Auto-publish | Drafts are queued and publish without a review step. You switch it on, per agent. |
| Manual | The agent runs only when you press the button, and its output stays in History. |
Auto-publish shows a warning-coloured caption while it is on, so you do not forget which agents have a long lead. Even then, a post that needs a picture or a video it does not have is held instead of being sent to fail.
Our advice is to lengthen the leash one agent at a time, after its drafts have needed no edits for a while. See human in the loop: when to approve and when to automate.
Things no agent can do
Some actions are not on any leash. They are behind a door only a signed-in person can open.
- Spend ad money. A boost starts when a person clicks Start in the web app. Agents, schedules, workflows and the assistant cannot prepare, start, resume or raise one.
- Message your followers. Auto-DMs send only a message you wrote, to people who comment a keyword you chose. The AI cannot write one, change one or switch one on.
- Send a job application. In Job Scout nothing is sent automatically, and there is no setting that turns that off.
- Send a follow-up. One is prepared; you send it.
- Connect an account or change billing.
These are enforced on the server. Hiding a button would not be enough.
A leash on cost, too
Autonomy can run away with money as well as words. You can cap AI spend for the whole account and for single agents. You get a warning at a threshold you choose, and at the cap AI work pauses before it spends more. A scheduled agent cannot drain your month.
Private stays private
Your Job Scout CV, application drafts and job-search details are not automatically used as sources for public posts. Brand Studio uses only what you add to it.
What the leash does not do
It does not make drafts true. A model can state something wrong with complete confidence, and the review step exists because of that. Read facts, names and numbers every time.
It also does not replace judgement about timing. Software does not know that today is the wrong day for a cheerful post.
The complete list is on control and safety.
Frequently asked questions
Can FinalPost AI publish without my approval?
Only if you switch a specific agent to Auto-publish, or set a docs source to auto mode. By default every draft waits for your approval.
What happens when the AI is unsure how to reply?
It escalates. The item is flagged as high risk, moved to the top of the inbox, and given a short holding reply instead of a guess, so a person can answer properly.
Can an agent spend money or message customers?
No. Paid boosts need a person to click Start, and automatic direct messages send only text you wrote and switched on yourself.