PROPOSAL: Agent can provide an input schema for client when the task state is input-required.

[Areo-Joe]

It is common for agents to require human-in-the-loop to gain more info/context/permission. And when this happens, structured input data is often need. Like a confirm option needs a boolean value, a select option needs user to choose from a list of things.

Thus, when the agent explicitly needs structured input, it can provide a input schema, which guides the client to provide info it needs.

Another reason for input schema/structured input is that the input is sometimes processed by a specific program first(code written by human), not directly handled by LLM/Agent. So It is necessary for this use case that ensures client passes correct data format for the agent server to run a code to process the input first, and then somehow trigger the underlying LLM/Agent.

With this added to the protocol, we may create a rather generic client UI that renders a form according to the input schema, and plug & play onto any agent server that may require a structured input.

[jpandrade30]

I know that A2A is not MCP, but are agents that need some complex information to take decision, and would be great to have some way to validate if the information being passed is enough or if something had changed in the mean time.