AI agents
An agent in Repost21 is an AI employee. It has a role (what it does), an instruction (how to behave), a “brain” (an AI model you choose yourself) and linked social media (where it works). One agent writes posts, another replies to clients in DMs, a third keeps an eye on competitors — each is busy with its own task and configured separately. All agents live in the “AI agents” section; you can view the list as cards or as a table.
Agent roles
| Role | What it does | Set |
|---|---|---|
| Content agent | Writes post copy from a brief — works in the composer | Basic |
| Inbox agent | Replies to clients in the Inbox on its own — using the knowledge base, with escalation to a human | Basic |
| Trend agent | Analyzes competitors’ posts, finds topics and formats, suggests ideas | Advanced |
| Outreach agent | Outbound campaigns — personalized messages to selected contacts | Advanced |
| Content autopilot | Autonomously generates and publishes posts on a schedule — see Autopilot | Advanced |
Basic agents (Content + Inbox) are the most in-demand — they’re created automatically and cover the two main tasks: making content and handling conversations. Advanced ones (Trend, Outreach, Autopilot) aren’t created on their own — you can install them from the template marketplace whenever you need them.
Template marketplace
Below your list of agents there’s a catalog of ready-made templates: preconfigured agents with a well-thought-out instruction and a set of tools. Installing takes a couple of clicks: pick a template, optionally set a name and social accounts — and the agent is ready to work. Popular templates show a “N teams use this” counter — social proof that the configuration works.
Catalog categories:
| Category | What’s inside |
|---|---|
| Basic | Content agent and Inbox agent — the starter set |
| Services · Education · E-commerce · B2B · Lifestyle · Support | Niche inbox agents: café, barbershop, fitness trainer, online school, online store, B2B SaaS, support FAQ bot and others |
| Content | Content agents for a specific task: personal brand, B2B expertise, local business |
| Advanced | Trend agents, Outreach campaigns, Content autopilot |
After installation a template is just a regular agent: you can change the instruction, model and links however you like.
Setting up an agent
Clicking an agent opens the editor. Main fields:
- Name and description — how the agent is labeled in the interface.
- Instruction (system prompt) — the key field: who the agent is, how it behaves, what’s allowed and what isn’t. Templates already have an instruction written — you can add the specifics of your brand to it.
- Model (“brain”) — which AI thinks for the agent: Claude, GPT, Gemini or DeepSeek. So you don’t have to figure out model names, there are human labels next to them: Fast (cheap, for simple tasks), Balanced (the optimum for most scenarios), Powerful (maximum quality for complex texts).
- Temperature — creativity: lower is more predictable and strict on facts, higher is more varied and bolder. For an inbox running on a knowledge base, low is better; for creative content, higher.
- Tools — what the agent can do: search the knowledge base, analyze competitors’ posts, escalate to a human, remember facts about a client, generate a branded image. The set depends on the role; knowledge base and brand library tools are connected automatically when you link them.
Linking social media
An agent is linked to specific connected accounts — the “Social media” tab in the editor:
- No links selected — the agent is universal: it works as a fallback for any team account.
- Accounts selected — the agent works only on them, and it takes priority over a universal one.
This is how an agency keeps a dedicated “Café” bot only on the café’s Telegram and a “Barbershop” bot only on the barbershop’s Instagram — plus one universal bot for everything else. Linked accounts are shown as chips right on the agent’s card.
Knowledge base and strict mode
Link a knowledge base to an agent — and it will answer based on your documents: prices, terms, FAQs. For an inbox agent with a knowledge base, strict mode is available: the agent answers only with facts from the base. Before each reply it searches for relevant fragments; if nothing is found for the question — it doesn’t make things up, but honestly tells the client it will pass the question to a colleague, and hands the conversation over to a human. This is protection against hallucinations: even a cheap model won’t be able to answer “off the top of its head”.
Brand library
Linking a brand library gives the agent the brand’s voice and visual style: texts respect the tone of voice, and generated images use the corporate colors, logos and references. A content agent with a brand library automatically gains a branded image generation tool.
Enabling and limits
Each agent has an on/off toggle. The number of simultaneously enabled agents depends on your plan — see the current numbers in the Plan section. You can keep as many disabled agents as you want: create drafts, experiment with instructions — the limit is only spent on enabling. If after switching plans you have more enabled agents than the limit, a warning appears — disable the extras or upgrade your plan.
Agent avatar
By default an agent has an avatar in its role’s color. You can upload your own image — a brand mascot, for example: the avatar is visible in the agent list, the sidebar and everywhere the agent is mentioned.
Agent analytics
The “Agent analytics” page shows the work of the whole AI team: how many actions were performed, how many tokens were spent and how much that cost in money — in total and per agent, over a period of 7/30/90 days, with a CSV export. Each agent also has its own “Analytics” tab in the editor: runs by day, statuses and a feed of recent actions. More on the metrics — in the Analytics section.
Cloning
The “Duplicate” button in the agent menu creates a copy with the same instruction, model, settings, knowledge base and brand library. Social media links aren’t copied — assign the clone your own accounts; it’s created disabled so you can check the settings first. A typical scenario: a second autopilot for a different social network or a variation of an inbox bot for a different brand.