Is Structured Human Input the Missing Link in Agentic Work?
Have you ever told your agent to go do something and walked away? You set the spec clearly. You defined what done should look like. You provided reference materials. You expected to come back later and find it finished. Yet, you come back and its paused waiting for your input. The times it needs input from you in unpredictable and uneven. This clearly does not scale. A long-running task can pause…
The article explores the challenges of human input in agentic work, specifically in the context of the Agent2Agent (A2A) protocol. When an agent is paused, it often requires input from a human user, but this input can be unpredictable and uneven, leading to inefficiencies. The article argues that a structured approach to human input is needed to address these issues.
A2A already has the right lifecycle shape for long-running work, with tasks that can be moved into an "input-required" state. The protocol can also carry structured JSON data in a "data" part, providing a starting point for handling human input. However, the article identifies a gap in the protocol: a standard way for the agent to describe the type of input it needs when it pauses.
The proposed solution is for the agent to provide an input schema for the client when it enters the "input-required" state. This schema could take the form of a boolean confirmation or a list of choices, depending on the nature of the input required. By sending a structured request, the client can render an appropriate control, validate the user's response, and send back structured data.
The article emphasizes the importance of treating human participation as part of the workflow state, rather than treating it as unstructured chat. This would enable operational questions to be addressed more easily, such as identifying tasks that are blocked on a decision or distinguishing between requests for approval and requests for missing account data.
With a portable input contract, clients can present the request in a way that makes sense for their specific UI, while preserving the same task identity and response shape.
The Agentic AI Foundation is bringing projects together around the protocols and infrastructure agents need to work across implementation boundaries. The A2A task model is an example of this work, demonstrating how agents can expose enough state for other systems to collaborate with them. The article concludes that moving human input with the task is a crucial step in building agentic applications that are flexible and scalable.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
