Google says Every company Will soon be on multi -agent systems -several AI agents who work together.
That sounds exciting and futuristic, but what I found in conversation with my customers is that, during agents sound Like something that we should all build, there is little information that helps you decide what your company should do.
Let me share an example of a complex company that is currently doing cool things with AI agents. Then we will deal with the agents and hopefully I will help you decide whether you should invest in this technology.
Walmart uses a number of AI agents. Each of them is designed for highly specific tasks. First, there is Sparky. At first glance Sparky looks like a normal chat bot.
But it does more than to chat. It connects with inventory, suggests products and can even suggest recipes and add the lack of ingredients to your car.
These relationships of AI -driven workflows make me excited about how agents change the way we do business.
Walmart has other AI agents, including one It was built for your employees to help you find things in business and deal with HR questions and other operational problems and a “Marty” other who works with suppliers and sellers to manage onboarding, orders and advertising campaigns.
The most interesting is the system of agents that Walmart uses to create digital twins. You create virtual 3D models of your shops and then use these digital twins to test things like storage layouts and carry out what-if scenarios. The Associate Agent could then receive a connection to real -time instructions for resuming with this agent. Or, and here is the part I think it’s wild. The digital twin of a business can find that a cooling unit will probably collapse. A message can then send to a repair technician in this shop, together with the required parts and an instruction manual about what needs to be done.
Walmart used Model context protocol (MCP) Standardize how many agents interact with the services they already have.
I can see the enormous value in what Walmart does, but how can we implement it in ideas that I can really implement in my own business or recommend it for my customers? Everything is so overwhelming!
Before we go on, let’s cover some basics.
Google’s definition of agents is that they are Software systems that use AI to pursue goals and do tasks. You can use argument, planning and memory and have autonomy that work, make decisions, learn and adapt.
A KI chatbot generally follows a script. An assistant uses AI to provide personalized help. An agent, on the other hand, can work proactively, use tools and even work with other agents to achieve complex goals.
Specialty | AI -agent | AI assistant | Simple AI chat bot |
---|---|---|---|
autonomy | High. Can operate independently to achieve a goal. | Medium. Supports users, but decisions are made by the user. | Low. Follows preprogrammed rules. |
complexity | Complex, multi -stage tasks and workflows. | Comes simple, single -tied tasks. | Automated basic, simple conversations. |
interaction | Proactive and goal -oriented. | Reactive, reacts to user inquiries. | Reactive, reacts to certain triggers. |
Learn | Can learn from experience and adapt his behavior. | Can only have limited learning functions. | Limited to no learning based on the interaction. |
I think of a AI agent like an employee. You can train them to do things and use tools. They give them a goal to achieve and they have an autonomy to determine the best way to do this. And if you are good, you will learn and improve.
Agents begin with an LLM – your agent can use chatt, gemini, claude, grok or other LLM. You can connect to tools via something called a model context protocol. Imagine MCP like a universal connection that makes it so that every LLM can work with a tool. The tool is essentially something that you could use program -controlled via an API. For example Google search, Google search console or possibly stripe. The difference between an MCP and an API is that the LLM with MCP does not need to know specific code Using for every single tool – it just has to describe what it wants to do with the tool.
In short, MCP enables the agents to speak to tools.
Communicating with tools is one thing, but what we really have to understand is what happens if Agents communicate with other agents. Google created a protocol for this – similar to a protocol for HTTP, mentioned A2A – agent to the agent.
Agent to agent communication
With A2A Every agent has a card This uses a simple language called JSON Describe what it can do.
Here is an example from Google:
The agent card could look like this:
# A2A Agent Skill definition skill = AgentSkill( Â Â id='get_exchange_rate', Â Â name="Currency Exchange Rates Tool", Â Â description='Helps with exchange values between various currencies', Â Â tags=('currency conversion', 'currency exchange'), Â Â examples=('What is exchange rate between USD and GBP?'), ) # A2A Agent Card definition agent_card = AgentCard( name="Currency Agent", description='Helps with exchange rates for currencies', url=f'http://{host}:{port}/', version='1.0.0', defaultInputModes=("text"), defaultOutputModes=("text"), capabilities=AgentCapabilities(streaming=True), skills=(skill), )
This agent could be part of a group of agents who can speak together.
In this example, when this travel agency speaks to a customer and realizes that you ask about booking a hotel, it can contact the hotel reservation agent and so on.
How are agents found?
In some cases, we only give our agents a list of agents with which they can communicate. In other cases, however, Google’s documentation indicates that the agent can come from catalogs or registers from agents.
Google has one Marketplace for agents. 945 agents are listed at the time of writing. However, it is not very easy for what I can see. Some of the agents have costs – but the documentation Simply says to connect with sales to get more information about pricing. I think we will at some point see catalogs of agents with whom we can connect. Or finally I assume that we can simply advertise you on our website and LLMS you will find.
Some agents are likely to appear in the search.
Google becomes agent
Glenn Genn has just posted that he Get the AI ​​mode for access to agent functions. It is currently quite primitive, but you can see which restaurants have reservations and what time.
These restaurants did not have to create agents to appear in AI mode. Rather, these restaurants used A Booking system from third -party providers This establishes a connection to something that Google has created with the title “Reserve with Google”.
At Google I/O we have seen the introduction of agents with which you can find tickets by connecting to websites such as ticketmaster and stubhub directly from the search results in AI mode. Google is looking for us and puts together information that helps us make our decision.
The AI ​​mode can then act as our agent and enable us to buy directly from the search results without having to visit a website.
We will not only find agents through AI mode, but also have the opportunity to act Gemini as our agent from a chrome browser. Google recently Announced Gemini in ChromeAnd say: “Get ready for your agent -Bowsing assistant.” They shared this example, in which our Gemini agent in Chrome reads a list of Google Mail and then goes to a website to put food items in a shopping cart.
As far as I can see, InstaCart does not use an agent here. Rather, Gemini uses Project Seepfahrer This is the AI ​​that uses its own Chrome browser to simply search the web and interact with websites.
I think most websites don’t have to do much to become part of the Agentic website, except that your website is easy to use of agents tools such as Project Mariner. This is something I would recommend now – use Chatgpt Agent and Project Mariner and watch how you do tasks on your website.
Agents will finally handle payments
Google announced a new protocol with which the agents entitled “Payments” can be treated Agent payments Protocol (AP2). They worked with Mastercard, American Express, Coinbase, Etsy, Salesforce and PayPal. This protocol will open doors for agents to safely meet the payments. This protocol is currently working with common payment methods, but is also set up with cryptocurrency. Also this month Stripe showed a new one Blockchain infrastructure called Tempo. And also presented chatgpt Agent purchases in cooperation with Shopify and Etsy.
In the development of active ingredients, this protocols creates the infrastructure required to transfer the money value. I think the day will come where you may hire my SEO agents for a contract where you pay me whether my agents have helped you earn money or improve your efficiency. I think the way our economy works will change significantly.
Will agents replace the search?
I personally believe that the operating use of the web will replace the traditional search. Here’s something The CEO of Google Deep Mind Demi Hassabis says…
“I think and so forth. If everything is based on agents, we will want our assistants and our agents to do a large part of the work and much of the everyday work that we are currently doing, right? You know, fill out forms, make payments, bay tables and something like that.
I think we will end with a kind of economic model in which agents speak to other agents and negotiate things between themselves and then give them back the results. And they also have the service providers with agents who offer services, and maybe there are some commandments, costs and such things and efficiency. Then I hope that from the user’s perspective, you have this assistant who is great that you … just like a brilliant human assistant and can take care of many of the secular things for you.
And I think if you follow this That means many changes to the structure of the web and the way we are currently using it…I find There will be incredible possibilities that appear economical and otherwise based on this change, but I think it will be a great disturbance. “”
What I have shared so far is a bit annoying, but let us continue to do things. It is one thing for Google to use our websites agal – so that Gemini use the web in Chrome or Chatgpt Agent and do things for us. But where things are really crazy, when we have agents who work with other agents.
I have a number of agents that help me get brainstorming for writing content. An agent thinks for ways to use my customer’s experience and create original, informative content. Another finds new research that we would like to write about for our audience. These are cool … but at the moment they are like individual employees who all work alone. What happens if we have a team of agents?
I may have an agent that establishes a connection to Google Search Console and GA4 with MCP. And this agent can send information to an agent whose task is to analyze pages that do well and find out which properties they have in common. And this agent would combine with a brainstorming agent that I can call when writing new content. Or let’s say I sell products. I could have agents who together with brainstorming about working what works to convert people into customers. While we find out how to do this, I could essentially work together with a goal of thousands of agents – to make my business more successful.
I learn to build these systems bit by bit. At some point the technology becomes easier and I believe that everything we can imagine can bring us to life.
My recommendations for companies at the moment
Here is what I would recommend.
- Spend every day with an LLM – I would stick to the acquaintances – Gemini, Chatgpt, possibly Claude and Grok. It is probably good to get used to the use of several voice models. Press your skills to try to help you do things in your work. Your goal is not to create a complete agent system that changes your company, but to learn the language of the conversation with these models. They are essentially just math, with the predictions based on the knowledge you have – and the more we learn to benefit from these tools, the better. If you have a company, I would recommend that every employee have the task of learning something new that he can do with an LLM every week.
- Create Gemini Dedelstones every day. These things are great. A Gemini Gem It’s just a request that you reuse regularly. It can have information in a knowledge base on which it bases. I use gemstones repeatedly in my day. I expect, but I don’t know that Google will ultimately make it easier for us to transform our gemstones into agents that can communicate. If you are a company with over 100 seats, you already have something like this about Agent space. As far as I can see, agent space can be expensive. The idea is that every employee can create an agent who works with company data. Personally, if I have a request that I use repeatedly, I make a jewel out of it.
- Consider learning the code. You have probably heard of the mood coding – coding with natural language. I do not encourage you to determine an app that you sell to the public, but instead learn to create something you can use. I made an RSS feed reader who is just great, and it saves me time. Then I learned how to combine this tool with my Google documents so that I can summarize stories and save in my newsletter. Little by little I learn how to encodes things. I think that those who can code have endless opportunity to create. I would encourage you to play with Openai’s Codex and also with The Build function of AI Studio. These tools both get better and do not make it possible to create things. If you learn now, you can build something at some point.
- Test your website with agent browsers such as Project Mariner and Chatgpt Agent. Make sure that your content is accessible to AI and that AI can fulfill things that a user wants, such as: B. fill in forms or put products in the car.
- Stay up to date on the use of agents in your area. I actually have Do not think that it is still time for my customers to build functioning agents That will change your business. Why? Because the technology changes quickly. I think every developer should play with Google Agent development kit And to see what you can do with it. And I think we should pay attention to every new news that helps us understand this technology more. I think Google will make it easier, but it will take some time.
This is probably a good place to mention my newsletter
Every day I spend hours reading and learning how AI changes our world so that I can share it with subscribers You can use AI messages.
New – my monthly report on the agent -Web
If you are a search bar member, Click here to get a 25% discount discode. If you are not a member of my community, Listen to!
We have so much to learn when it comes to agents. At the moment I think we should all learn and be careful and understand how to build and use AI agents. I think as soon as people earn real money with agents, we will suddenly take off this field. Let’s see!
Marie