Autonomous Mode: The Two Controls That Keep It Safe

Autonomous mode is safe to turn on because two controls sit in front of it, and both are real code, not settings theatre. The worker counts the posts your AI has already made today and stops at your plan's cap. The approval gate is on by default, so nothing reaches a public account until a person releases it. One control is not enforced yet: the monthly spend cap.

On a schedule, a background worker picks up each brand you have enabled. It reads that brand's Brand Brain — voice, audience, reading level, sentence length, always-say facts, banned words, emoji policy, hashtag ceiling, CTA style, claims policy — and generates from those constraints. It writes the caption, scores hashtags, writes the CTA, generates an image where you want one, and slots the post per platform and timezone at that platform's native dimensions and character limit.

Then it stops. The post exists in your queue. It is not published. Publication is a separate step, and by default that step needs a human.

That sequence matters for trust: the model is conditioned before it writes, not corrected afterwards. If a post comes out wrong, the fix is usually a field in Brand Brain, not a rewrite of that one caption. More on the mechanics in how SocialPower works end to end.

Every plan carries a maximum number of autonomous posts per day. Before the worker creates a post for a brand, it counts the AI posts already made for that brand today. If the count has reached the cap, it does not generate. It does not queue the overflow for tomorrow. It does not run and then discard. It stops.

Because the limit is a count of real posts rather than a number held in memory, it cannot drift. Posts you write yourself are not AI posts, so your own scheduling is unaffected by the cap and does not eat into it.

The cap is the ceiling, not the target. On Pro you may run at 5 a day; most brands are better served by 1 or 2 that survive review. Full limits are on the pricing page.

The approval gate ships switched on for every workspace. An autonomous post lands in your queue as something to release, and you see exactly what would go out: the caption at that platform's character limit, the hashtags with their competition scores, the CTA, the image at native dimensions, and the slot it is booked into. You approve it, edit it, reschedule it, or delete it.

This is the control that makes the rest of the system reviewable rather than merely fast. It also decides who can act. Team roles run owner, admin, manager, editor, viewer — a client on viewer access can watch the queue fill and never be able to push anything live.

There is a second reason the gate matters. SocialPower supports OpenAI, Claude, Gemini, Groq, Mistral, DeepSeek and HuggingFace, set per workspace, with a deterministic mock provider as a fallback so a bad key or a missing config never hard-fails the app. That fallback is a stability feature, not a writing feature. Mock output is placeholder text. With the gate on, you see it and stop it. With the gate off, you would not.

Be clear about this one. The monthly spend cap is stored and displayed in the interface, but the worker does not enforce it. Reaching that number does not halt generation. Treat it as a reading, not a brake.

Your actual cost controls today are two, and neither of them is that field:

• The daily post cap, which bounds how many generation runs happen per brand per day. This is the one the worker obeys.
• A hard billing limit set in your AI provider's own dashboard. Since keys are per workspace, set the ceiling there — OpenAI, Anthropic, Google or whoever you have configured. Do this before you enable autonomous mode, not after the first invoice.

Spend-cap enforcement is tracked on the roadmap alongside the other work in progress.

Do not go from zero to unattended. Each stage has an exit condition you can actually check.

1. 01Weeks 1-2: approval gate on, cap at 1 a day. Review every post before release. Count how many you edit. Every edit is information — if you keep deleting the same phrase, add it to banned words; if the CTA is always wrong, fix CTA style; if a claim is too strong, tighten the claims policy. You are training the record, not the model.
2. 02Week 3 onwards: gate off, cap held at 1 a day. Move here only if you edited fewer than one post in five during week two. Holding at 1 means the worst case for a bad week is seven posts you have to clean up, not fifty. Read the queue daily anyway — you are watching output you no longer block.
3. 03Week 5 onwards: raise the cap, one step at a time. Go 1 to 2, sit for a week, then 2 to 3. If the edit rate climbs again, drop back a step. On Pro, most single brands settle between 2 and 4 a day; the fifth slot tends to be filler.

If you are still rewriting most captions in week two, the problem is not the cap and not the model. It is an underspecified Brand Brain — usually missing always-say facts and an unset reading level. Fix that before you touch any autonomy setting.

Agencies should run this per client, not per workspace. Each brand carries its own Brand Brain and its own connected accounts, and nothing crosses between them, so a client who is fine with unattended posting and a client who insists on approving every word can sit side by side.

Worth stating plainly, because the word "autonomous" invites assumptions. Five agents are on the roadmap and labelled SOON in the product: trend detection, SEO, rendered video, analytics readback, and competitor tracking. Autonomous mode does not use them, because they do not exist yet.

It also does not touch comments, DMs or paid ads. There is no comment moderation, no inbox management, no ad spend. Nothing in autonomous mode replies to anyone. It publishes to Instagram, Facebook, LinkedIn, X, Pinterest, Threads, TikTok and YouTube — see the platform-by-platform detail — and that is the whole of its reach.

And because there is no analytics readback yet, autonomous mode does not learn from performance. It does not notice that Tuesday reels do better. It generates against your Brand Brain, consistently, on a schedule. Consistency is the thing it gives you. Optimisation is still your job.

Ten minutes of setup removes most of the risk. In order:

1. 01Fill banned words with the phrases you would never publish — competitor names, overclaiming words, anything legal has flagged.
2. 02Set the hashtag ceiling and the emoji policy explicitly. Left blank, they are the two settings most likely to produce a post that reads as machine-written.
3. 03Write three always-say facts: what you sell, who it is for, and one thing that is true and specific — a price, a city, a delivery window.
4. 04Set the claims policy to the strictest version you can live with. Loosening it later is a five-second edit; retracting a published claim is not.
5. 05Confirm the approval gate is on, set the daily cap to 1, and connect one account for one brand — not all of them.
6. 06Open your AI provider's billing dashboard and set a hard monthly limit in rupees. SocialPower will show you the spend, but it will not stop you.