Investigate User Input / User Approvals

[westey-m]

We need an ADR to explore how an agent can ask the user for approval of unsafe operations, or how an agent can ask the user for structured input.
Various agentic protocols support these, but we also want to ensure that anyone building a custom agent, can use these patterns.
Some inference services are also supporting approvals now, so this so we should also consider how this affects IChatClient implementations.

[ekzhu]

For user input within an agent, shall we consider this to be a response from the agent? When human provide feedback / approval, it re-runs the agent with a specific input message type or perhaps a chat option.

Making the input / approval process external to the agent has many benefits. For example, it allows a workflow involving the agent to be suspended and persisted while waiting for the input to arrive -- it often takes much longer time for human to write something than machine.

[westey-m]

@ekzhu, thanks for the input. Yes, I also think this approach makes the most sense. Check out the ADR document in full and please add your thoughts there as well: #209

[ekzhu]

Follow up with workflow design for HIL: #399