Original Summary

Building the MCP server took a weekend. Making it survive real agents took months, and the interesting part is that none of the failures were protocol problems. They were product problems that agents expose faster than humans do. Here is the list. The agent does not know which account you meant. The moment a brand has two Facebook Pages or two LinkedIn orgs, "post this to LinkedIn" is ambiguous. Every post now has to carry one account id, and the tool description tells the agent to list accounts first and check that the account is still connected. Disconnected accounts were already the number one cause of failed posts for humans. Agents made it worse because they do not notice the red badge. Validate at create time, not at fire time. A scheduled post fails hours after the agent finished talking, and nobody is there to read the error. So media is required at create time on the platforms that need it (TikTok, YouTube, Instagram, Pinterest), even for drafts, the media type has to match the file, and size caps are enforced on upload. The agent gets the rejection while it can still fix it. Retries duplicate. Our create call has no idempotency key. If the transport drops and the agent retries, you get two scheduled posts. Listing and deleting posts are tools too, so the agent can clean up, but that is the honest gap and it is next on the list. Do not make the agent wait for the publish. The MCP call only creates the scheduled post; publishing runs later as a server job. A 200MB video never sits on the MCP transport, which is where every timeout report we saw came from. Approvals need to be a tool, not a dashboard. Posts can be created as pending approval, and either a person or the agent, via an approve call, moves them to approved. Nothing publishes while pending. Without this, teams turn the agent off after the first wrong post. Some platform limits should be enforced by you, not discovered by the agent. X charges far more for posts with links, so we cap link posts per month per plan. An agent that loves adding links hits the cap on purpose. Assistants quote your docs verbatim. When someone asks ChatGPT for a scheduler with an MCP server, the answer lists the vendors whose pages print the endpoint URL and the setup steps in plain text. Pages that describe the feature in marketing language are skipped. We rewrote ours accordingly, and it is the cheapest change on this list. Live-data features are hard to get through app-store review. Our ChatGPT app listing was rejected twice because the review environment could not reproduce inbox test cases: comment fixtures decay, and a token with a smaller scope 403s every inbox tool. We resubmitted with the inbox left out of the listing, not out of the product. If your feature depends on live third-party data, plan the reviewer's fixtures before you plan the launch post. Questions for anyone exposing a write API to agents: how are you handling idempotency on create, and do you let the agent approve

  • 情报分类：工作与职业机会
  • 分类依据：内容涉及招聘、求职或职业发展
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/16 20:21:00