#713: AWS News: Meet the Next Generation of Amazon SageMaker, Multi-Agent Collaboration on Bedrock

0:00This is episode 713 of the AWS podcast released on March 24th, 2025. five. Hello everyone and welcome back to the AWS podcast. I'm going to share with you. Great to have you back. I'm joined by one of my two co-hosts. I'm joined by Gillian Ford. G'day, Gillian. G'day, Simon. It's always exciting to be here on the AWS podcast. It's always fun. And Shruti is off gallivanting at some fancy schmancy conference. So we're hoping she hears us say that. So she'll be embarrassed that she didn't join us for the podcast, but she's doing some interesting research for the podcast while she's away. So she'll share that with us.

0:35Now, a few things before we get crackalackin. Firstly, again, we want to do a mailbag episode. We've got a few really interesting questions that have come in, but we want more. AWSpodcast.amazon.com is the place to send it. If you've got a question about AWS, about Amazon, about the services, about the podcast, it's up to you. Throw them in there, and we look forward to answering those questions. But let's have a look at some of the key things that have happened in the last couple of weeks. Firstly, SageMaker Unified Studio has gone GA. Now, this was announced a little while back, but it's basically a unified studio for data and AI development together.

1:12And I guess, Gillian, like I'm not a data guy per se. I don't have that deep data engineering expertise, but people that do tell me that this is a big deal. Like to have it all in the one place saves a lot of time. What are you hearing? Yeah, I think so too. And I think now the delineation between data engineering and AI is really just starting to become just one unified, all in the same team here. And I think that's what you're starting to see with SageMaker Unified Studio. Yeah, because you've got all together. You've got Athena, EMR, Glue, Redshift. You've got Bedrock in there as well, so you can build and customize your apps.

1:50Q Develop is in there as well, so that's for the workflow. There's a SQL editor right in there. You know, my fanboyness of SQL. So lots of stuff happening there. Plus there's also S3 tables is also continuing to be integrated with things like SageMaker Lakehouse, which you can query straight from there as well. So there's a bunch of stuff that I think is going to make things easier. And it'll be interesting to, I guess, see what people do, because also you can publish your data and your AR models to the catalog as well. So you can get secure governance. You can use Amazon data zone capabilities.

2:26so you can make sure that the right people get the right access to the data. It's not just about putting it together. It's who has the need to know as well. Yeah, I'm really excited to see how customers start to use SageMaker Unified Studio. Interesting times. And speaking of interesting times, Amazon Bedrock now supports multi-agent collaboration in GA. Now, I think this is a big deal. And I say that because I've been doing some work lately with an application that relies upon multiple agents. And basically what this is, is networks of agents with supervisors and specialized sub agents that can each do their own thing and do them well.

3:01And I think this is interesting, Gillian, because, you know, if we try, what we're learning is if we try and make a single AI call or an AI, even a single LM do a whole bunch of stuff at once, they struggle, but they're really good at doing chunked up jobs. And this makes the chunking even easier, doesn't it? It does. And I think it's cool that it now allows you to think about picking the right model based on what the task is and being able to cost optimize because maybe you want a smaller model that will be cost less, maybe be faster. And then maybe where there's something that requires more thinking, more intelligence, you can use a more capable model in that sense.

3:45And there's a number of ways that these interactions can happen. So firstly, there's supervisor mode and supervisor with routing mode. And with routing mode, the supervisor agent will just route simple requests directly to the subagent. So I won't do the full orchestration dance, but in supervisor mode, it can do amazing things automatically across them all. But the other thing I like is that you have inline agents, which means you can dynamically adjust the roles at runtime, which I find fascinating. I think this is going to be really interesting as the supervisor determines emergent requirements of the workflow.

4:17it can then change what the roles are doing to adjust. And also the fact that you can reference payloads outside, so external data sources without having them embedded in the request. That's another big thing. I mean, there's so much here, it's almost difficult to unpack in one go. It is, yeah. I mean, there's just so many different types of use cases that I feel like this can really unlock in any kind of industry. Like, I definitely see it a lot in, like, healthcare, for example. A lot of these complex workflows of being able to triage patient requests, for example, are coming up. Anything that comes to mind that you're seeing with this?

4:55Yeah, look, I think I'm really interested in the concept of engineering and design and the ability for multiple agents to perform a complicated team-based task in a synthesized way that then has human oversight, which kind of means that you have access to a team. So for example, it could be a testing team or it could be a design team. It could be an ideas team, what have you. And each of those team members will have different capabilities. And this lets you sort of bring them all together in a really specialized way, which is just, it's just very interesting. And I think there's been a lot of talk about sort of agentic AI.

5:32I think it's multi-agentic AI that's actually the interesting part. But time will tell. But certainly in the work that I'm doing, I'm seeing it be really, really useful, which is fun. Yep, I totally agree. Let's get on to some of the other updates that have happened. Firstly, analytics. We're happy to announce support for AWS Data Glue catalog views with AWS Glue 5.0. So this allows you to create views from your Spark jobs that you can query from multiple engines without requiring access to reference tables. Again, security is important. Amazon Kinesis Data Streams now supports IPv6. So if you need support for that, you've got it.

6:07And Amazon EMR on EKS now supports Amazon EKS pod identity. So this reduces complexity and possibilities of errors. Some great updates for application integration. AppSync Events adds publishing over WebSocket for real-time PubSub. Where was this a month ago when I was doing PubSub stuff? I could have used it then. But starting now, you can use your AppSync Events APIs to publish events directly over WebSocket connections implement, which complements the existing HTTP API publishing capability. So this allows applications to both publish and subscribe to events using a single WebSocket connection, streamlining the implementation of real-time features.

6:48This is a win. The other thing that's really interesting is the announcement of Adibus Step Functions Workflow Studio for the VS Code IDE. My hand up, I'm a VS Code kind of guy. I use it all the time. And now you can visually create, edit, and debug your state machine worklines directly in your local development environment, which is cool. Amazon Bedrock now supports fine-tuning for Meta's Llama 3.2 models. These are for the 1B, 3B, 11B, and 90B, enabling businesses to customize these generative AI models with their own data. Fine-tuning allows you to adapt Llama 3.2 models for domain-specific tasks, which allows you to enhance performance for specialized use cases.

7:34Amazon Bedrock KnowledgeBases now supports GraphRag, while it's actually generally available. So GraphRag delivers more comprehensive, relevant, and explainable responses by leveraging relationships within your data, improving how generative AI applications retrieve and synthesize information. I'm really excited about this because I'm starting to see a lot of customers that are really interested in utilizing graphs, and especially those who haven't used graphs before, but see the applications within generative AI to support a variety of use cases. So if you're new to graph, I definitely suggest going back in the time machine.

8:152024, we did an episode on graphs. We talked about Amazon Neptune as well. So you'll definitely want to scroll back into the AWS podcast feed to find that episode. And this is, so this has got Amazon Neptune like built in, isn't it? So it's like you don't have to do anything. There's no barrier. Let's just get into it. And what's cool is that GraphRag, so it automatically is going to generate and store vector embeddings in Amazon Neptune Analytics, along with a graph representation of entities and their relationships. GraphRack combines vector similarity search with graph traversal enabling higher accuracy when retrieving information from disparate yet interconnected data sources.

8:59Also check out the AWS blog on this one because we did some really, and a really interesting blog post that showed you the difference between using GraphRack versus not using GraphRack. And it's just a great way of being able to see the subtle difference of what the graph algorithms are able to provide. Amazon Nova Pro Foundation model now supports latency-optimized inference in preview on Amazon Bedrock, enabling faster response times and improved responsiveness for generative AI applications. Latency-optimized inference speeds up response times for latency-sensitive applications, improving the end-year-zure experience and giving developers more flexibility to optimize performance for their use case.

9:40Amazon S3 Tables now seamlessly integrates with Amazon SageMaker Lakehouse, making it easy to query and join S3 tables with data in S3 data lakes, Amazon Redshift data warehouses, and third-party data sources. Amazon Bedrock's capabilities are now generally available within Amazon SageMaker Unified Studio, offering a governed collaborative environment. DeepSeq R1 is now available as a fully managed serverless model in Amazon Bedrock. AWS is the first cloud service provider to deliver DeepSeq R1 as a fully managed, generally available model. You can power your applications with DeepSeq R1's capabilities through Amazon Bedrock's fully managed service via a single API, along with Bedrock's tools, allowing your team to focus on building differentiated generative AI applications right away.

10:30DeepSeq R1 combines powerful natural language processing with advanced reasoning capabilities demonstrated by its leading performance in mathematics were achieved 79.28 % on AIME, I guess that's AIME, maybe that's how it's pronounced, 2024, and software engineering where it achieved a 49.2 % score on SWE Bench Verified. DeepSeq R1 provides a cost-effective option for unlocking advanced AI reasoning capabilities efficiently, maximizing your technology investment. Amazon SageMaker Inference now supports rolling updates for inference component endpoints. Amazon QDeveloper announces a new CLI agent within the command line.

11:14You can now ask QDeveloper to write code, test it, help debug issues, and QDeveloper will iteratively make adjustments based on your feedback and approval. Amazon QBusiness now supports insights from audio and video data. SageMaker HyperPod flexible training plans now supports instant start times and multiple offers. Amazon Lex now supports confirmation and alphanumeric slot types in Korean locale. AWS Health Omics now supports the latest NVIDIA L4 and L40S graphical processing units and larger compute options of up to 192 vCPUs for workflows. Next topic is business applications. AWS Wicker is excited to announce a redesigned admin experience that's now fully integrated with the AWS Management Console.

12:08Let's talk about compute. Amazon ECR has announced ECR to ECR pull through cache. So this is a capability that allows customers to automatically sync container images between two ECR private registries existing across AWS regions and or accounts. So this means that you can get the reduced latency of pulling those cached images in region. And so this makes it more cost effective, faster, and more reliable as well. Amazon ECS now offers GPU optimized AMI for Amazon Linux 2023. And Amazon EKS now enveloping crypts all Kubernetes API data by default. So this is an additional layer of security with an AWS-owned KMS encryption key or the option of bringing your own key.

12:54Bottlerocket now supports NVIDIA multi-instance GPU or MIG for Kubernetes workloads. So this allows you to maximize your GPU resource utilization by running multiple workloads simultaneously on a single GPU whilst maintaining hardware level isolation between those workloads. Bottlerocket also now supports adabus neuron accelerated instance types this now allows you to deploy and manage machine learning inference and training workloads on a bunch of new instance types like inf1 inf2 train1 and train2 so you can move faster and bottlerocket also simplifies system setup with default bootstrap container image so you have something to start from rather than having to maintain your own container image for that initial configuration now let's talk about customer engagement amazon connect has broadened global telephony coverage there is now industry-leading inbound number availability across 158 countries and national outbound numbers in 72 countries so this expansion increases coverage by an average of 125 across regions telephony is hard we're trying to make it easier amazon connect contact lens can now dynamically update the questions on an evaluation form which is pretty nifty and amazon connect amazon workspaces and amazon appstream 2.0 and now Chrome Enterprise recommended.

14:13So this ensures seamless integration and performance for businesses using Chrome devices. And Amazon Connect can now also target multiple agent proficiencies in a single routing step. So you can have up to three all conditions and you can match a contact with four different types of agents and increase the possibility of finding a suitable match. So for example, if the backup for a niche banking skill consists of agents trained on account management, registration, and tax. And so what this means is that you can match the right agent to the right customer at the right time with a higher probability of success.

14:47Let's talk about databases. Amazon Aurora PostgreSQL zero ETL integration with Amazon Redshift now supports multiple integrations. In fact, you can have up to five from the same Aurora cluster. So this makes it even easier for you to move data loads across without having to do a lot of work. Amazon RDS for SQL Server supports new minor version in February 2025, and Amazon RDS for MySQL has got extended support for minor versions as well. My regular reminder for you to patch your stuff. Now let's talk about developer tools. AWS Code Connections now allows you to securely share your connection resource across individual AWS accounts or within your AWS organization.

15:31Previously to create a connection, you had to install the AWS Connector app for GitHub or GitLab or Bitbucket for each AWS account from which source access was required. But now using AWS Resource Access Manager to securely share a connection to your third-party source provider across AWS accounts. By using AWS Resource Access Manager to share your connection resource, you no longer need to create a connection in each AWS account. AWS Code Connections now provides greater control to manage the creation of hosts with a new IAM condition key for self-managed GitLab GitHub Enterprise Server hosts.

16:11AWS CodePipeline introduces CodeBuild and commands rule for stage-level condition. AWS CodePipeline supports invoking pipeline execution with a new action type. AWS CodeBuild now supports registering self-hosted runners at GitHub organization or enterprise level. Additionally, you can assign your self-hosted runners to specific runner groups for enhanced security and access control. AWS CodeBuild is a fully managed continuous integration service that compiles source code, runs tests, and produces software packages ready for deployment. Aduus makes it easier for developers to build serverless applications by bringing serverless LAN's extensive application pattern library directly into the Visual Studio Code IDE.

17:00This integration eliminates the need to switch between your development environment and external resources when building serverless architectures by enabling you to browse, search, and implement pre-built serverless patterns directly in Visual Studio Code IDE. Yay, IDEs. We love them. Let's do an update, a quick but important one for end user computing. Amazon Workspace's pools now supports FIPS 140-2 validated endpoints. This is a US government standard that specifies some security requirements for sensitive information. So this one's important. And let's move on to the topic of front-end web and mobile.

17:39AWS Amplify hosting announces deployment SKU protection support. Now, this feature ensures front-end requests are always routed to the correct server backend version, which eliminates what's called version skew and makes your deployments more reliable. You can enable this at the branch level, and there's no additional cost, my favorite price, and it's available to all customers who already use this particular service. It's a great way to manage version changes through your user base. This one's exciting because I don't think we've had a game tech announcement on the news update show in a while.

18:12It's been a while. It's been a while. Yeah. Well, for those who have been waiting anxiously, Amazon GameLift Servers launches GameServer Wrapper to rapid onboarding. Developers told us they face significant overhead integrating the server SDK, including dependency management and custom code implementation. The GameServer Wrapper solves these challenges by eliminating the need for server SDK integration, making it easy to deploy game servers on Amazon GameLift servers with zero code changes. It's annoying. I'm invoking my right to make dad jokes on the podcast. So this is not MC Game Lift rapper.

18:49This is a code rapper. So I didn't know what that meant. I was making a very bad rap pustache. MC Game Lift. Maybe that's why. I don't get it. But I succeeded. And we've got another one from Amazon Game Lift Streams. So Amazon GameLift Streams is a new managed capability that allows developers to stream games at up to 180p resolution and 60 frames per second to any device with a WebRTC enabled browser. In just a few clicks, you can upload games built with a variety of 3D engines with little to no modification, provision streaming capacity in specific AWS regions, and immediately start test streaming.

19:38players can start playing no yeah this is really cool players can start playing triple a double a and indie games over the internet in just a few seconds on their pcs phones tablets and smart tvs without waiting hours for a download let's talk about the internet of things announcing mqtt enabled sitewise edge gateways for adbus it sitewise so of course adbus it sitewise is a managed service that makes it easy for you to collect, store, organize, and analyze your data from industrial equipment at scale. With this launch, newly created gateways now include an MQTT v5 broker component that centralizes connectivity and makes it easier for you to get up and going.

20:21One quick update for management and governance, Amazon EC2 allowed AMIs now integrates with AWS Config. So you can now use your AWS Config rules to automatically monitor, detect and report instances launched using AMIs that have not been allowed by the allowed AMIs rule. So this means that you can make sure everyone's doing what they should be doing. Now let's talk about networking and content delivery. Amazon Route 53 Traffic Flow introduces a new visual editor to improve DNS policy editing. AWS Application Load Balancer now allows customers to provide a pool of public IPv4 addresses for IP address assignment to load balancer nodes.

21:01Customers can configure a public IP address manager pool that can consist of either bring your own IP addresses that is customer owned or a contiguous IPv4 address block provided by Amazon. Let's talk about security, identity, and compliance. Amazon S3 Access Grants has simplified authentication when using both IAM and identity provider permissions. So it just makes life a little bit easier. You don't have to choose between them. It'll figure it out. Amazon Inspector has expanded ECR support for minimal container-based images and enhanced detection. So this means that scanning goes quicker and you can maintain better security with very robust environments.

21:43Amazon EventBridge has expanded IAM execution role support to all targets. So this means you can have consistent permission policies and dedicated invocation throttling limits. adibus waft now supports pci dss 4.0 compliance protection with partner solutions and adibus waft adds ja4 fingerprinting and aggregation on ja3 and ja4 fingerprints for rate-based rules so this is really important for you folks who work in that difficult world of figuring out who's who and what's going on and who's good and who's not so good and how to protect yourself against that. Amazon Redshift Data API now supports single sign-on with AWS IAM Identity Center.

22:27IAM Access Analyzer now supports IPv6. You can pick the trend. We've been talking about IPv6 for a while. I'm pretty sure pretty much everything will be IPv6 this year. AWS Service Reference Information now supports resources and condition keys. And onto the topic of storage, Amazon Data Firehose now delivers real-time streaming data into Amazon S3 tables. So if you're a Firehose user, you can Firehose it right into the tables. And Amazon S3 has reduced pricing for S3 object tagging by 35%. There you go, another cost reduction. You didn't have to do anything for, but you get it. And Amazon S3 tables adds Apache Iceberg REST catalog APIs.

23:06So any iceberg compatible application can easily create update list and delete tables in an S3 table bucket. And Amazon S3 tables has added create and query table support in the S3 console. So if you don't have a tool, you can just jump in there and get it done. Adibus Backup has added logically air-gapped vault support for Amazon FSX. So a logically air-gapped vault is a type of Adibus Backup Vault that allows secure sharing of backups across accounts and organizations, supporting direct restore to reduce recovery time from a data loss event. A logically air-gapped vault stores immutable backup copies that are locked by default and isolated with encryption using AWS-owned keys.

23:47And finally, Amazon FSx for NetApp on tap no longer charges for SnapLock licensing, so it's even more cost-effective for you to protect your business applications. There's a bunch there today, Gillian. There was a really interesting mix of infrastructure, a lot of storage stuff. There was some really interesting developer stuff, like something for everyone. Really, yeah. And even the price reduction, I mean, it's been a couple of episodes since we had one, and especially, I think, of 35%. So that's really exciting. Yeah, it's meaningful. How do folks reach out to you? Julian Ford on LinkedIn. Excellent.

24:20And AWSpodcast.amazon.com is the way also. And of course, until next time, keep on building.