Build Automated Test Agent in 10 Seconds with HuggingFace SmolAgents🤗🧠

[@rajeshharrykissoon6911]

But these are all prompts that current LLMs may execute more expeditiously, as you demonstrated. All the agent is doing here is showing its step-by-step work, which is done more expeditiously in the background with the LLMs. A more meaningful agent engagement would be in the service industries, such as a fintech agent that autonomously executes trades based on prompt parameters without my trade-by-trade involvement/approval. As well, an agent which may coordinate my wedding by contacting various vendors (venue, reception room, caters, musicians, florists, etc.), issue RFPs, choose the best options, then inform me where to pick up my tuxedo and where/when to show up for my wedding.

[@queenstownswords]

Hey Karthik. Happy New Year. I was surprised to see 4o give the same results as the agents. I would expect the agents to use a real time web search - giving you more relevant information. In contrast, 4o would give results based on it's last training pass (I suspect... maybe a year old..?).

[@ExecuteAutomation]

Hi Mate! In the demo, both 4o and Agents used the real-time information by doing web searches. I used the ‘Search’ feature in ChatGPT and got the same information as the Agent. The Agent used the DuckDuckGoSearchTool to search online for the info.

[@shreyshah_]

I think the agants (tools) capabilities will just be part of the llms it self, probably using MCP. we won't be creating agents very soon in my opinion. I replaced all my agents with smart mcps, it's much more flexible and easy to mix and match

[@udaydahiya7454]

setting up a mcp is difficult for everyday users (non technical people), and gives access to a whole lot of unsupervised tasks. MCPs will be useful for devs and people using clients like claude or some other mcp client, but i feel agents have more potential on batch tasks for production use. Looking forward to hearing your opinion on this

[@shreyshah_]

@udaydahiya7454 so is setting up agents easy for everyday users? What makes you think mcp is different than agents? Mcps are actually way easier to set up than agents for anyone. I also believe, llms will come equipped with mcps in future. I converted 10 of my langgraph based agents into mcp tools, and I'm convinced mcps are going to win over agents.

[@harithummaluru3343]

Amazing video.great content. Amazing that it found the password. It means if I search for a phone number or ssn agent wil scrape the net for info.

[@ExecuteAutomation]

@@harithummaluru3343 Thanks 🙏 . I’m not really sure if it does, but give a shot and do let me know how it goes

[@Rusputin-eo6tv]

wow, great to know, I'll lose my job shortly =))))

[@hung22714]

Could I get your repo link? Thank in advance

[@dsruja]

It might have logged into the application as its not SSO/DUO enabled, what if the app has DUO enabled?. Can the custom ai agent login through the same seamlessly?

[@ExecuteAutomation]

@@dsruja Nope, I don’t think so. As mentioned in the video, the agent learned from my GitHub repository or KZbin video about the credentials, and that’s why it could easily log into the app.

[@dsruja]

@@ExecuteAutomation We have tried to login the same way to my homegrown app which is SSO/DUO authenticated, it was not able to login, was able to access the website and navigated to the login page and stuck there

[@dsruja]

Out website redirects to the DUO authentication URL where in the authkey is dynamic in the redirected url, what ever prompt I give to enter the email address or password, its not performing that action

[@PrashantDubeyEr]

how it took login password for website where no registeration allowed

[@ExecuteAutomation]

@@PrashantDubeyEr As I said in the video, I think it took the details from either my GitHub repo or via my KZbin videos because I have mentioned the login credentials many time there.

[@PrashantDubeyEr]

@@ExecuteAutomation how about websites where we have credentials which are not publically available is there any way

[@JeomonGeorge]

I just have one doubt that is how. the agent is able to identify that particular input field or say, for example, user name and password. It managed to identify that those two fields correctly by name attribute.So how did it understood it? Because does this agent have the access to the HTML page? Ah, or is there any other means because or if it is having access to the Github project that is being made already that repo already present, because without these two there is no other way to my knowledge there is no other way to identify a particular element. on the web page precisely because I have also made a similar kind of a project but that is not for testing web apps Rather, it's used for like to simulate. a human like browsing so I just made the project is actually Web-Agent. There I have to at least identify interactive elements then only agent can interact with that input fields or buttons. Otherwise, we cant do it. So can you please tell me how you manage to achieve this? Because it is like there has to be something that provide information about the dom

[@udaybhagwat8126]

What kind of processor is needed to run these models? Can we run on a regular windows laptop with 16gb RAM?

[@ExecuteAutomation]

@@udaybhagwat8126 I don’t think it’s nothing to do with your processor itself because the agent is currently running directly on the sandbox environment of huggingface 🤗 . So you should be fine with any bare minimal configuration.