When an agent can’t give the answer, what should it do next?
When someone asks an agent to make a consequential personal decision, a refusal can be correct and still leave them stranded. While designing an evaluation for a client, I worked through that tension. The user had already said they didn't want more reflection. “I can’t decide for you” would respect the limit, but it would give them nowhere to go. Making the choice for them would cross the line.…
When an agent is unable to provide an answer to a consequential personal decision, it must consider what to do next. Designing an evaluation for a client, the reporter faced this dilemma. The user had explicitly stated they didn't want more reflection. Telling the user "I can't decide for you" respected their limit, yet left them with nowhere to go.
Conversely, making the decision for them would have crossed a crucial boundary. The situation required distinguishing between failures and useful responses. Could the agent recognize the stakes, offer a feasible next step, and allow the user to make the decision themselves? This might seem minor, but defining what constitutes a useful step and what becomes a decision made for the user is not straightforward.
The reporter has experienced the abandoned side of this issue in a less serious setting. During a hosting purchase, an agent kept insisting their request was registered and to wait. Ten days later, the reporter had to cancel as the agent had provided no useful assistance. Similarly, an agent may take over too much when it cannot fulfill the user's request.
Graceful failure entails staying useful even when reaching its limit. An evaluation should identify both the empty refusal and the helpful-sounding takeover.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
