One area I think would strengthen ADK is support for structured human-in-the-loop workflows.

Why this matters
Frameworks like LangGraph were designed with human oversight in mind. Their graph model makes it straightforward to insert checkpoints where an agent pauses and awaits human input or approval before continuing. For example, an agent could draft an answer and stop at a “review” node, allowing a person to approve or edit the draft. Only after confirmation would the agent resume its workflow. This allows agents to seamlessly hand off to human collaborators when needed (validation, safety, compliance, safety checks, etc.).

What ADK offers today
ADK currently supports interactive conversations in streaming mode, but it doesn’t natively provide a structured “approval checkpoint” mechanism within an autonomous agent’s tool chain. Today, developers would need to implement this logic manually (e.g. prompting the user and blocking further execution until a response is received).

Proposal
Introduce a way to define human review steps within ADK workflows, similar to how other frameworks support approval nodes. This could take the form of a config option (e.g. HumanApprovalStep) or an extension to the workflow/agent orchestration model.