Відео

Hey. Have you seen this video? ua-cam.com/video/C_AtgCm43Nk/v-deo.htmlsi=R8HTCMR0Ibqq2ck0

@mikegchambers yes actually I have. And I got a different perspective of performing sklearn based deployment from this video. Problem is I am trying to work on tensorflow deployment model but on numerical data and somehow I have tried everything but during training job my model is running but its not getting saved and i am getting a warning. I am new to ml.I have tried like everything to the best of my knowledge. Even refereed documentation. And tensorflow based model deployment but on numerical data is something I am after.

Well, in the video I show it working with Llama 3.2, so yes, no problems.

@@mikegchambers thanks😃

Awesome. Thanks for comment.

Hey. Thanks for trying out the code. You will need… an AWS account with credentials in your dev environment. You will also need to have enabled access to the LLMs models, through the Bedrock console page, clicking model access near the bottom of the menu on the left. If you have any issues, reach out and I’ll help you get it sorted.

@@mikegchambers I think I fixed the problem, thanks a lot!

In theory yes. But you’ll need to modify the code quite a bit. Also you need a model that can handle agentic workflows. This might be a little challenging for a quantised smaller model running on most “home” hardware. Please let me know if you make progress.

Hey. So that library is not out yet. I was talking to the team at Anthropic today and it seem they want to get a good authentication design working before they do that. However, there is a workaround, which is to make a local stdio server and have it make a connection across the web. But that is a hack of a workaround only. Let’s keep an eye out for updates.

Glad you found it interesting. Thanks for the feedback Jim!

Glad you found it useful. I can see me making a bunch more on this.

Yeah. Sounds about right. I need to dig in to the code more. From my memory it’s just updating an in memory string ‘memo’. I’m not exactly sure what the point of it is, there is no tool to query it, just write. :/

Okay. Had a long chat with the developer at Anthropic who wrote the SQLite server. There is an asynchronous call from the server that is looking for my response out of band of the agentic workflow. I’m going to work on this and get a fix. I’m also going to work on the other parts of MCP like prompts and sampling. It was s super useful and great chat.

@@mikegchambers ❤️

Hey! Well… it’s a protocol right? So it’s as fast or slow as the implementation. I was please to see that my super basic code implementation is pretty fast for basic DB operations.

@mikegchambers i see got it, thanks man, noted.

Absolutely, go for it! I think I for deploying larger solutions in production we can often benefit from an API hosted model, but yeah, if you have the hardware that would be fun.

Well… okay. Glad you enjoyed the beard. 🧔 😆

Yay! We have a winner! Thanks. I was beginning to think it was too obscure. You win… my respect!

Thanks for watching!

Thanks!!

I refuse to believe that Anthropic didn’t realise this connection!!

Hi. Thanks for your feedback. I’d be interested to dive a little deeper I think. Agents are certainly more complex than some other frameworks, and I’d say that they are different from most other frameworks for a reason. Bedrock Agents (aside from this announcement) are designed for production scale deployments of robust and simple agents. Whenever I do anything more complex I weave Bedrock in to another framework etc. The other point I grapple with is that we need a functioning AWS environment to use these agents. As such we need to set up all the security etc, and as this is probably where the rest of the app stack lives, I need to pay careful attention to getting this just as I need. So it’s not the case that one API key can simply take care of everything as I’m doing a lot lot more. Keen to know more about what you think. (If you don’t mind I’ll make more videos!! Haha.)

@@mikegchambers Here's my rant: An agent, at its core, is a construct designed to reason, act, and adapt in a dynamic environment. The boundaries of an agent include the following primary components: reasoning mechanisms, action execution, state observation, and a feedback loop to iterate and improve decisions. Auxiliary components - like additional integrations, orchestration, or monitoring-enhance its utility but are not strictly necessary for its core functionality. Bedrock, at its essence, acts as a proxy layer - particularly when paired with models like Claude under the hood. The majority of the other supported models, such as Titan and AI21, lag significantly behind industry leaders in terms of performance and capability. For Bedrock to truly shine, it needs to align itself more closely with the expectations of robust agentic frameworks. When considering deployment, containerization and horizontal scaling are key to production-grade agents. However, Bedrock currently falls short in enabling this easily. The abstraction Bedrock offers adds layers of complexity without providing the flexibility to containerize and scale seamlessly. In contrast, SageMaker, while not perfect, is a more mature service in terms of offering well-thought-out integration and deployment strategies for ML models. The problem isn't just defining agents - although Bedrock could provide clearer, better-defined mechanisms for reasoning, execution, and feedback loops. The bigger issue lies in orchestrating communications and interactions between agents or sub-systems. Bedrock supports rudimentary protocols that feel slow, bloated, and lack the polish required for a good developer experience. Calling an LLM or chaining API calls does not make a system agentic. True agentic behavior involves carefully structured reasoning processes, well-defined action protocols, and a robust feedback loop - all of which Bedrock handles loosely, often leaving developers to grapple with the missing pieces. AWS's pitch that Bedrock simplifies workflows feels like an overpromise. In reality, it adds unnecessary complexity and feels like features are being released for the sake of hype rather than solving real developer pain points. As a long-time AWS user (15+ years), it's frustrating to see an attempt at abstraction that ends up introducing more hurdles rather than making life easier. This isn’t to say Bedrock has no potential - it does. But for now, it feels like a half-baked solution that needs a more focused approach to simplify and enhance the agent development experience. Let’s hope AWS takes user feedback seriously and evolves the platform into something truly production-grade. (This response was refined using ChatGPT btw)

Thanks for the detail there. If you’re at reinvent we should meet up and discuss. I’m taking in a lot of what you say. One reply I have, and this is not addressing, or designed to address the majority of your points, but on the deployment side, and scale… one of the points of Bedrock Agents is that the scaling is managed for you. So if you use Lamda for the logic you don’t need to think about containers etc. Let me know if you want to meet at reinvent or jump on a call sometime. Thanks - Mike

Glad it was helpful!

Glad it was helpful!

@@mikegchambers They always are, just like your Sagemaker course!

So kind!

Thanks!

You’re welcome! 🎉

Haha. Hopefully I’ve gotten better at videos over here years. (Probably not :) )

For that architecture you would want to place it behind an API. Typically you would use the API Gateway and a Lambda function.

Yeah. Where I landed, for various reasons, was a Cloudformation template with custom resources. So I guess this could be flipped to other IAC.

Welcome! :)

Thanks. Thinking about doing more of this kinda stuff.