Over the past year, the word “agent” has landed on the homepage of almost every social media tool. It’s used to sell three very different things: a text generator with a “write post” button, a chatbot running a decision tree, and a system that decides on its own what to do and when. From the outside, all three are “AI for social media.” Inside, the gap is roughly the one between a calculator, an answering machine and a new hire on probation.
The difference isn’t academic. If a generator writes nonsense, you simply don’t publish it. If an agent replies to a customer at 3 a.m. with a price that doesn’t exist, that price is now your problem. In 2024 a Canadian tribunal ordered Air Canada to compensate a passenger after the airline’s chatbot promised a fare discount its actual policy didn’t offer. The airline’s argument that the chatbot was responsible for its own words didn’t fly.
Below: how to tell an agent from a bot, five levels of autonomy, which social media tasks to give an agent right now, a step-by-step setup, and where agents reliably go wrong.
Generator, bot, assistant, agent: what’s the difference
| Tool | What it does | Who decides | Typical task |
|---|---|---|---|
| Generator | Writes text on request | You, at every step | ”Write a post about our fall sale” |
| Scripted chatbot | Walks users through a button tree | Whoever wrote the script | ”Choose: 1 — prices, 2 — address” |
| AI assistant | Completes a task you ask for, using tools | You set each task | ”Draft three posts for next week and schedule them” |
| AI agent | Chooses its own actions toward a goal, within limits | You set the limits and approve | ”Every morning, check the sources and prepare two drafts” |
An agent is a model plus tools (calendar, publishing, analytics, knowledge base), a goal and a loop: observe → decide → act → check. If a system does nothing without your command, it isn’t an agent, whatever the landing page says. That’s fine — for many tasks an assistant is more than enough.
Five levels of autonomy
The easiest mental model is the one used for self-driving cars — levels:
- Suggestions. You write; AI offers hook or headline options. Zero risk.
- Assistant with tools. On request, it drafts a post, adapts it per platform and schedules it. Every action is triggered by you.
- Agent with approval on every step. It works on its own schedule, but every output lands in an approval queue.
- Agent with approval on exceptions. Routine work — digests, recurring series, answers to common questions — goes out on its own; exceptions go to a human.
- Full autopilot. No human in the loop. In social media it’s justified only for narrow tasks, like announcing each new blog post in your Telegram channel.
Move up one level at a time. A simple rule: if for two weeks straight you edit fewer than one draft in five, you’re ready for the next level. If you’re editing every other draft, the agent lacks context, and moving up will only make mistakes faster.
Social media tasks to hand an agent today
Monitoring your niche and drafting. The agent reads RSS feeds, public Telegram channels and websites, picks what’s fresh and prepares 2–3 drafts a day. You choose instead of inventing. The sweet spot is 8–15 sources: fewer and the drafts get repetitive, more and the agent drowns in noise.
Per-platform adaptation. Mechanical work an agent does more carefully than a person, because it never forgets the limits: 2,200 characters for an Instagram caption, 4,096 for a Telegram post, 1,024 for a Telegram media caption, 500 on Threads, 280 on X without a subscription.
Calendar slotting. Spread 10 finished posts across the week’s best time slots without putting two posts from the same series back to back.
Answers to repeat questions. Opening hours, address, prices, availability, shipping terms — strictly from the knowledge base, with anything outside it handed to a human.
A weekly summary. “What grew, what dropped, which post brought inquiries” — one page, built on the numbers from our guide to social media metrics.
What not to hand over, even at level 4: replies to complaints and negative comments, reactions to news and incidents, collaboration negotiations, and any promise of discounts, deadlines or compensation.
Setting up an agent, step by step
- One task, one platform. Not “run our social media” but “prepare two drafts a day for our Telegram channel from these eight sources.” A narrow task is easy to evaluate.
- A knowledge base instead of a giant prompt. Pricing, FAQ, shipping terms, off-limits topics — in live documents that update without you. A price that isn’t in the base must never show up in a post.
- Voice. 3–5 reference posts and a list of banned phrases — the method is in our post on brand tone of voice.
- Explicit rules. What’s allowed, what isn’t, and when to call a human. A working instruction looks like this:
Every weekday by 9 a.m., prepare two drafts for the Telegram channel based on the connected sources. Up to 900 characters, one idea, end with a question to the reader. Not allowed: prices or deadlines not in the knowledge base; news about accidents or incidents; mentions of competitors. Summarize foreign-language sources in your own words and link the original. When in doubt, don’t publish — ask.
- Approval in a messenger. For the first two weeks, every action needs sign-off. A Telegram notification with “Publish / Edit / Reject” buttons takes about ten seconds per post and doesn’t require opening a laptop.
- An action log. What the agent did, when, and based on what. Without a log you’ll never figure out where an error in a post came from.
- A quality metric. The share of drafts published without edits and the number of human handoffs per week. That’s your evidence for moving up a level.
Where agents reliably go wrong
- Made-up facts. A model will confidently fill in a price, address or deadline if it isn’t in the context. The only cure is a hard rule: facts come from the knowledge base; everything else is a question for a human.
- Tone drift. After a month, posts get smoother and more alike. Once a week, read five random posts in a row — drift jumps out immediately.
- Blindness to context. A cheerful scheduled post on the day of a local tragedy, or in the middle of your own outage. You need a “pause everything” button, and whoever is on duty needs access to it, not just the admin.
- Loops. The agent replies to a bot, the bot replies back, and a 40-message thread grows under your post. Cap the number of replies per thread.
- Gray-area actions. Mass liking, mass following and commenting under other people’s posts break platform rules. An agent should work only through official APIs and only with your own content. Even there, limits apply: Instagram allows 50 API-published posts per rolling 24 hours.
When you don’t need an agent
- Three posts a week on one network. A calendar and scheduled posting are enough; setting up an agent will cost more time than it saves. We covered where that line sits in SMM automation.
- No knowledge base. An agent without facts is a generator of confident fiction.
- Nobody to approve. In the first weeks an agent needs 15–20 minutes of attention a day. If you don’t have that, start with an assistant.
- Regulated niches. Healthcare, finance, legal services: drafts only, no autonomous replies to customers.
- The brand is the founder. If people follow you for a specific person, that voice can be sped up, not fully delegated.
In Repost21, agents work from your knowledge base in Google Docs, Sheets or Notion, and you pick the mode per task: approve every post, approve only exceptions, or autopilot for routine. Luna, the AI assistant, sends the approval queue straight to Telegram. Setting up your first agent is covered in the docs, and you can try it on the free plan.