Read-Only AI Agents
Posted on:
The current hype train around Generative AI is around Agents: programs that can take action on your behalf. They are promised to go browse the web, read through your data and knowledgebases and take "intelligent" work against what you ask it. "Book me a flight to LA in October" for some reason is one of the things Google, OpenAI, Anthropic, and the rest of the companies always gives an example for agents. That said out of all of this, I think what I would like at the end of the day even if these agents stop hallucinating and actually become useful is leave the "action" to me.
Let's run through that book flight example I just brought up. One could imagine it will go searching across websites for flights to a certain location. It may ask for certain criteria like:
- Do you care about long layovers or do you wish for more direct routes?
- Do you want early departures or don't care about when you leave?
- Do you put in the starting airport as a constraint or are you willing to drive to another nearby airport to get a better flight?
- Do you have travel points / memberships with particular airlines?
And then with all of that information give me a listing of flight itineraries. Then if you decide on a particular flight it could go and book it for you. That last part is actually where I would want an agent to give me the reins again. I would allow it to aggregate everything for me but then I get to pick what flight I want and actually book it. Even if the hallucination / error rate of these agents went to 0% and I could have 100% certainty it would do it correctly, I still feel I would want to do the actual action at hand.
And outside of the example of booking flights, this still stands that I would prefer agents to be read-only. In another example, say there was an agent sitting in a meeting and captured the entire conversation and gave a summary and action items at the end, yes I could have it throw issues up or add things to my todo list, but I feel doing that myself as well is not losing too much time. And as well, who is to say that an agent would have the nuance to know what I actually need to do. Say person A tells me to do something they think would be beneficial but the rest of a team doesn't. Is the AI going to be able to pick out that background context or is it just gonna add a task cause person A said so.
Across the board, I feel as humans we shouldn't give full control away to AI for either things that seem mundane at this current state. These tools may be able to get better over time, but I feel there is some joy of saying "I did this" and actually having understanding of the action that was done rather than leave all of the heavy work to some program.