How to improve multi-step MCP call execution with instant models

Since GPT-5 and later models possess excellent agentic capabilities, I am implementing a workflow where MCP calls are coordinated effectively, culminating in a UI call.

This behavior works very well with thinking models; however, with instant models, they sometimes return a response without completing the sequence of consecutive MCP calls. Because of this, I am struggling with how to write the MCP descriptions.

I suspect that providing instructions for the entire MCP application or implementing a feature for Skill loading would resolve this to some extent. Is there any plan to introduce such features?

Since the Codex Plugin already provides Skills and MCPs as a set, it would be ideal if they could be provided in a similar manner.

1 Like