Exploring Google’s Agent Builder: A Comprehensive Review

Google has recently made waves in the tech community with its announcement of the new AI Agent Builder during their latest conference. The emphasis was clear as the term "agent" was mentioned 46 times in a single keynote. This tool promises to facilitate the creation and deployment of enterprise-ready generative AI experiences using natural language or code interactions. But does it live up to its ambitious claims? Let’s dive into a detailed evaluation of Google’s Agent Builder and see if it’s worth your consideration.

Navigating to Agent Builder: A User’s First Challenge

Reaching Google’s Agent Builder is not the most straightforward journey. It starts by navigating through Google Cloud, where you first need to log in. The process involves heading to the apps section and creating a new application. For users located outside the US, the initial attempt to select a global region might stall, as certain functionalities appear limited to the US. This could be a significant deterrent for global users needing accessibility.

Setting Up and Testing the Agents

Once logged in, you’re faced with options to build and deploy agents. Google’s interface allows users to input instructions and set goals for the agents. For testing purposes, a ‘YouTube Title Writer’ agent was created, aiming to generate engaging and click-worthy video titles. The expectation was to interact with a simple user interface where on one side you set up the agent and on the other, you test its responses.

Selection of Language Models

A concerning aspect noticed during the setup was the limited choice of outdated language models. Even with the announcement of the new Gemini 1.5 Pro model, users were only offered models like Palm 2 or Lambda, raising questions about the rationale behind not incorporating the latest and potentially more powerful models.

Interface and User Experience

The Agent Builder interface, while aiming for functionality, comes across as complex and possibly overwhelming to new users. One needs to continuously save progress, and options like the ‘Code Interpreter’ aren’t universally available, depending on your selected region. The necessity of dealing with terms like ‘token limits’ hints at underlying limitations in the design, considering newer models should ideally handle larger data sets more efficiently.

Practical Testing and Results

The practical test involved generating several title variations for a hypothetical video. Unfortunately, the initial results were underwhelming with the agent either misunderstanding the requests or providing irrelevant answers. Even attempting to improve outcomes by introducing a second ‘Idea Analyst’ agent to critique video ideas didn’t smooth out the process. Responses remained disconnected from the set expectations, leading to frustration.

Multi-Agent Interaction

Attempting to utilize multiple agents demonstrated another layer of complexity and potential confusion. Commands to interact between agents did not execute as intended, and critiques lacked depth, often citing a need for internet access which wasn’t available.

Conclusion: Enterprise-Ready?

After an extensive trial, the performance of Google’s Agent Builder was disappointing. The tool seems to struggle with user-friendliness and efficacy, raising significant doubts about its readiness for enterprise use, especially when alternatives like CreAI, Lang Graph, and Autogen promise more user-centric and efficient solutions.

This firsthand experience might reflect teething issues that could potentially improve over time. However, for businesses and developers seeking reliable and sophisticated AI agent building tools today, exploring other platforms may prove more beneficial. For those interested in delving deeper into AI agent development, consider seeking communities and workshops that provide robust educational content and support, tailored to creating effective AI agents.

Google’s Path Forward

Google has the resources to refine and enhance the Agent Builder. For future iterations, focusing on user experience simplification, expanding access to updated models, and ensuring global functionality without regional restrictions could significantly impact its adoption rate and overall user satisfaction.

In conclusion, while the anticipation for Google’s Agent Builder was high, it appears that for now, its practical application and ease of use fall short of expectations, especially compared to competitors actively enhancing user experiences and broadening capabilities.

[h3]Watch this video for the full details:[/h3]


If you’re serious about AI, and want to learn how to build Agents, join my community: https://www.skool.com/new-society

Follow me on Twitter – https://x.com/DavidOndrej1

Please Subscribe.

Google just released their new Agent Builder, here are my honest thoughts.

[h3]Transcript[/h3]
so Google just held a conference where they announced their AI agent builder in fact they mentioned the word agent 46 times 46 times in a single keynote that’s kind of crazy so let’s take a look if it’s any good Google agent Builder just to get to agent Builder is quite a challenge so we have to go to Google Cloud okay seems like I’ve logged in successfully so let’s see build and deploy Enterprise ready generative AI experiences create AI agents using natural language or code for approach seems like they’re really aiming this to big companies you know let’s try it console. cloud. gooogle these URLs are getting crazy okay so you have to go to apps and then create a new app search Chat recommendations and agent in the bottom right corner select by the way this is crazy because for the region when I tried Global because I’m not in the US it didn’t let me create an agent all that worked was us so let’s try Global maybe it works now okay it’s promising by the way look at this URL Tex AIC conversation. cloud.google.com like this looks like a fishing link all right so let’s see let’s do like um YouTube title writer all right now on the right as you can see this is the preview so on the left this is kind of similar to when they’re building a GPT on the left you’re building it on the right you’re testing it so on the right we can select our agent and then the llm now this is very disappointing these are the old llms like I don’t know if it’s Palm 2 or even Lambda like these are super bad so I have no clue why they are recommending it and I have no clue why they are not adding Gemini 1.5 Pro here which is also what they announced at this conference today we’re taking Gemini 1.5 Pro into public preview but let’s go with the best model here Gemini 1.0 okay so YouTube title writer default goal so let’s see okay so we set it goal and instruction so that’s how they broken up the system prompt all right so the goal of this agent is obviously interesting and clickable titles okay instructions and actually Google gives us some pre-filled text so let’s see greet the user then ask how you can help summarize the user’s request okay oh this is interesting we can give it a tool and an agent I guess we can build multi-agent teams like this so I’ll definitely try this so we can have like a someone who writes the titles and someone who analyzes the video idea so let’s see always write 15 different title variations use simple and plain English okay and then after you finish writing all titles okay let’s see wait if I do dollar sign agent and then agent name so we don’t have a second agent created so I cannot do that okay let we use agent to critique the video idea and then we we’re going to fill this in afterwards okay so let’s save it again because we have to save all the time so on the left you have agents tools conversation history and settings but inside of building the agent we have Basics which is you know name goal instructions and then tools which only they have code interpreter and I don’t know why we are getting these errors oh yeah because I cannot okay you cannot use that without an agent actually there is a code interpreter I mean we can enable it but it’s code interpreter is not supported in your apps location Global let’s go to examples right no examples created yet each agent should have one or more examples remember this very good prompting tip these are effectively F shot prompts yeah exactly so let’s create an example right let’s do reading AI books this is one of videos I want to make what is this selection strategy but a token limit why are they talking about token limit I guess yeah this is the issue they’re using Gemini 1.0 because Gemini 1.5 pro has 1 million tokens so in that case they should never talk about token limit being reached let’s just do always okay and then we can do display names and we can do create titles and language English obviously summary of the agent execution result okay okay so I’m just going to say here are your title variations so I just included examples that I personally like if the agent can produce similar titles that would be great agent State what is this man okay canceled I mean I don’t want my agents getting cancel that’s for sure add an action manually agent utterance who in their right mind would choose the word utterance Google has too many employees this is so unintuitive it’s crazy provide an an utterance for this conversation okay it generated something whatever I’m going to delete this 15 new titles for this video idea okay so it completely ignored my request the video idea is about me reading every book and like we have one title two okay two out of 15 were good I don’t know let’s just leave this empty probably because I don’t want that to mess things up so now our agent has at least five examples let’s test out our agent okay so let’s cancel this and let let’s select the best model and our agent and let’s test it out help me write new titles for this video idea ranking free ey tools sure I can help you with that what is the video idea is this what everybody was panicking about Google agents are going to take over the world are you sure about that you know what let me let me do something let’s load up a terrible llm let’s do Gemma 2 billion I’m going to give it the same prompt in fact I’m not going to even include the examples okay so this is 2 billion a really bad really bad language model and I’m going to give this prompt and look at this look at the difference Google agents is this all you can do I already told you the video idea okay let’s give it again okay now it works but these titles are nothing like the examples I gave it what is this man this really feels like talking to gpt2 let’s add our second agent maybe it will get better once we introduce multiple agents right so let’s go back and let’s add add a new agent and this time going to be it’s going to be idea analyst okay objectively analyze the video idea provided by the user avoid politically correct answers at all cost whatever just three metrics that came to mind let’s save it again so okay we have our idea analyst saved let’s go back uh YouTube title writer uh okay and then we need to insert the agent right so dollar sign agent ID analyst okay and then save so this is pretty clear let’s make it even clearer always call and then let’s do agent ID analyst make it Ultra clear that way there is no excuse if it fails right so let’s reset this reset conversation so now we have a blank conversation again I have no clue why they’re offering these ancient models I mean help me write titles for my next video enter let’s take it you know one step at a time what is your video about my video is about Sam ultman making open AI open source again so we have the titles but it did not call the second agent I mean okay maybe I can see some influence like here is why this could be influenced by our example if you go to examples okay I have here is how so maybe it is influenced a bit I have no clue honestly let’s maybe say like now use the idea analyst agent to critique the video idea what happened now use the okay what idea analyst so it did call the second agent I’m sorry I can’t critique this video idea I’m not able to access the internet to get the context of the video okay I’m just going to say you don’t need any context just critique the idea itself I’m sorry I can’t critique the video idea without any context I need to know what the video is about this is wild guys the video is about Sam timan making I’m sorry I can’t critique the video idea this is terrible by the way just to prove that I’m not nitpicking this is what my friend was able to achieve with Google agent Builder I don’t like potatoes right okay I will create a meal meal plan for you that does not include potatoes here is a healthy and diverse meal plan for one day taking into account your preferences lunch salmon with baked potatoes another example I understand you want to avoid potatoes and fish is there anything else no that’s all and then it creates the diet look at this look at this here is your personalized diet plan breakfast food snack food lunch food how is Google expecting any of their Enterprise clients to use this Enterprise ready by the way guys this is Enterprise ready you know so if you have a company with thousand employees this is for you I feel like I’m talking to like a mental ill person you don’t have to see the video it’s not made yet I understand however I can’t critique the video idea without seeing the video itself I mean I don’t know if it’s really this bad nobody will use this and definitely not Enterprise clients I mean everybody will use crew aai Lang graph autogen there is no point in using Google agent Builder like if you want to actually build working agents then use crew or autogen studio and if you want to learn how to do that then check out my six module Workshop on how to build and deploy AI agents which is available inside of my community so if you want to be among people who are building AI agents we have hundreds of different agent Builders and we have materials in- depth step-by-step tutorials that are showing you how to build agents that actually work if you’re interested in that then make sure to join it’s the first link in the description