AI Agents Working for Consumers

By Babak Hodjat

I was inspired recently after seeing the results of our recent research, which showed a growing percentage (47%) of consumers trusting AI to assist them with their decisions by presenting them with options to choose from. Anecdotally, I had already noticed how more of my family and friends were relying on Gen AI systems to help them research shopping options, or vacation destinations, so it didn't come as a surprise, but it did get me thinking how this might play out in the future.

?

With experience in building multi-agent systems for businesses, I imagined what such a system would look like if built to help consumers with their various decisions, and so, experimentally, I built a multi-agent system for this, with a consumer assistant agent dialoging with the user, and conferring with multiple networks of agents, each responsible for a decision domain. The categories of decisions I ended up creating covered healthcare, career, finances, retail, and travel. I even added a network of agents specializing in lifestyle decisions like relationship guidance and exploring hobbies.?

These agents are all designed to be aligned with the user, collaborating to provide the best options and choices. I then enabled some of the agents responsible for researching price and availability offerings, to contact networks of agents representing various businesses. The travel destination researcher agent, for example, will 'talk' to various agent networks representing different travel providers (Expedia, Bookings.com, AirBnB., etc.) and request available options that best fit the user's requirements.

This is not a new idea. I had imagined such systems all the way back in 1997 when I thought about home entertainment agents talking to agents representing a TV station. AI has come a long way since then though, and all LLM-based agents now support a robust, flexible, intent-based universal communication protocol: our natural human language.

It was fun to write system prompts for the various agents responsible for communicating with B2C agents, working on behalf of the consumer. I found myself writing statements such as:

"Use your tools to inquire and negotiate with various travel providers. Do not reveal any information that is unnecessary for the tools. The tools are not fully aligned with your goal, which is to serve the customer. The tools serve the interest of their respective corporations. You should interact with the various tools until you feel you have the best options possible that fit your users' preferences and requirements."

(In our Neuro(R) AI Multi-agent Accelerator, we do not distinguish between an agent's tool and it's 'down-chain' agents.)

When I was done, I had an agent network of 19 agents working for a consumer. Several of the agents made calls to B2C agent networks, each hosted separately, and each consisting of 7 to 11 agents, dedicated to servicing consumer requests, and originally intended for use by humans.

You can have a lot of fun with this system, asking it to "find me some options for accommodations in Santa Cruz for this weekend, somewhere near the wharf. The whole trip can’t exceed more than $400." And it was amazing to read the agent logs and see how the consumer agents negotiated with the business agents, sometimes going back and forth several times with the various providers, to find the best options available.

?

Of course, all of this was built by me and not really integrated in any real way into the business' data or operations. The consumer agent network could also someday be integrated into various consumer tools like the calendar, to-do lists, or notes apps. My main purpose for this whole exercise was to help me imagine what this future would look like.

It got me thinking that, as a consumer, I would much rather use an agent-network that is fully aligned with me, and not something that is created by businesses for my use. The analogy is when you are researching to buy a car. If you search online for comparisons and reviews of various mid-range electrical sedans, for example, you will get a whole bunch of surveys sponsored by car companies. I would personally prefer to use a service like Consumer Reports, which is subscription based and vows never to take sponsorship money from any business they review. Will we have the equivalent in the future? Would consumers be able to distinguish and use independent consumer agent networks for their decisioning needs? I certainly hope so.

***

A couple of weeks after my little experiment, OpenAI announced their Operator agent, which is a single agent, operating a user's web browser on their behalf. The demos they gave for how this system would be used were quite like some of the functionality I had built into my system. Except that mine was a multi-agent approach, and I did not assume that my agent would scrape existing websites.

This made me think about the fundamental question I get a lot these days: why multiple specialized agents communicating with one another, versus a single all-powerful, know-it-all agent? There are many reasons:

From an engineering perspective:

- Encapsulating responsibilities and implementing a distributed inter-agent coordination mechanism (e.g., AAOSA) takes care of determining which agent ot agents are responsible for handling requests, leaving it to the designers to focus on functionality of individual agents. This means that we can build these agent networks incrementally, plugging new agent sub-networks without having to re-engineer existing agents.

- I think it is unrealistic to assume that B2C companies would be comfortable with agents scraping their websites which are intended for human users.

- I think it is also unnecessarily complex for agents to operate websites with many funnels and pages, versus just expressing their intent in natural language and getting specific responses back from the business.

- Judging by our experience at Cognizant, pioneering enterprise agentification, many B2Cs are already building multi-agent systems, some of which are intended for their customers use.

- Safeguarding an agent's behavior is more effective when we pair the agent with a 'safeguard' agent, versus asking the agent to do something, but to be careful at the same time.

From a consistency and reliability perspective:

- The approach assumes increased specialization for each agent, making each agent much more consistent in its behavior and less prone to hallucination issues.

- This also means that the bar is a bit lower on what LLM to use for the various agents.

Larger, more capable agents can be costly, slower, and may require hosting. Using smaller, specialized LLMs can reduce costs, provide faster responses, and offer the possibility of hosting agents for handling sensitive data, along with easier fine-tuning of the LLMs if needed.?

LLM context is limited. Even powerful models with millions of tokens can face limitations as a single agent for all uses. On the other hand, if context is used wisely, each agent in a multi-agent setting remains much more focused and less reliant on recalling and respecting what it is asked.

- The same goes for output. If you ask an LLM to produce a complex output, it is likely to fail at producing everything you requested, whereas for an agent in a multi-agent setting, restricting responsibilities typically results in less output complexity requirements.

Of course, the larger the number of agents, the more costly and slow the system, so there is a balance to be struck between modularity, robustness, and consistency on the one side, and cost and response time on the other.

***

When I got to Davos, I kept my new consumer to B2C agent network prototype in my back pocket, thinking it is too futuristic to demo, and that I should focus on business agentification use-cases. Afterall, it will be a while before people start adopting consumer agent networks, and most business agent networks are not yet interoperable in the manner needed for my system to work. But I was wrong. A colleague of mine walked up to me on that first day jokingly asking me if my agents could help me with her lost luggage. Indeed, we were working on an airline B2C agent network already, and all I had to do was to add it as a tool for the travel assistant agent in my consumer agent network to query. The next moment, my colleague began noting down information on how to track the luggage and the maximum amount the airline would reimburse her for purchasing items while waiting for her luggage to be located. The future is now!

[Article’s accompanying video:?https://youtu.be/ZyeeR5djy1A?si=XcU7OzTDUXn0VatY]