Short answer: let the agent draft and fill, never publish. Give it access through a connection you can limit and revoke, keep the final yes with a person, and make sure every change it makes is labelled as the agent's.
What can go wrong
- The agent publishes a draft you meant to review.
- It approves a post on the client's behalf, or skips the client altogether.
- It cancels or moves posts you had already agreed with a client.
- It fills ten slots on the same account in one afternoon, and the account looks automated.
- You can't tell afterwards which changes were yours and which were the agent's.
A safe setup, step by step
- Connect with sign-in, not a pasted key. Use a connection where you sign in and approve access, so you never paste a password or long-lived key into a chat. Standards such as theModel Context Protocol (MCP) let agents like Claude and ChatGPT connect to tools this way.
- Start read-only. Let the agent read the calendar and suggest posts for a week before you give it write access.
- Limit the scope. One workspace or one client at a time, so a mistake can't spread across every brand.
- Keep publishing with people. The agent may fill a slot; a person decides that the post is final. If your tool can't separate "filled" from "final", don't connect an agent to it.
- Keep client approval in the loop. The agent can send posts to the client. It should never be able to answer for the client.
- Apply the same posting limits. Pacing rules should apply to agent-filled posts exactly as to yours.
- Review the history. Check which changes came from the agent for the first few weeks, then spot check.
Good first requests
- "List every open slot next week for this client and draft a post for each, matching the slot's format."
- "Show me which accounts have less than 7 days of posts planned."
- "Send next week's filled posts to the client for approval."
How Suarga handles it
Suarga has no built-in AI. Your own agent connects to the Suarga MCP server and signs in with OAuth, and you choose read-only or read and write access, and which workspace it may use. Agents can fill slots, apply templates, create campaigns and send posts for approval. They cannot lock a slot, and only locked posts are published. They cannot approve on a client's behalf, apply or lift a freeze, or cancel posts. These rules are enforced by the scheduler itself, every agent change is labelled in the history, and pacing applies to agent posts too. MCP is included on every plan and during the trial.
Read what agents can and can't do, theMCP reference, and how agents fill slot holds you planned in advance. Setup guides for Claude andChatGPT are on the agents page.