The week of 6 July was the one where AWS migration tooling stopped assuming a human is clicking through it. DMS Schema Conversion became drivable from Kiro, Claude Code, and Cursor through the AWS MCP Server, and AWS Transform for migrations added an expert mode that takes your inputs up front and runs the steps without the chat back-and-forth. Around them, Security Hub started probing your estate from the public internet instead of inferring reachability, Aurora DSQL change data capture went GA, and the GPU management fee on EKS Auto Mode and ECS Managed Instances fell by up to 60%, backdated to 1 July.

Schema conversion becomes an API your agent calls

DMS Schema Conversion has had a generative AI assist for code objects since re:Invent 2024. What changed on 10 July is the control surface. An agent connected to the AWS MCP Server can now create a conversion project, browse source metadata, run the conversion, generate the assessment report, and export the result, from natural-language instructions in the IDE. A dms-schema-conversion skill loads on demand and carries the API ordering, exclusion patterns, and operational sequencing so the agent does not have to rediscover them. It covers every existing source and target engine pair and adds no charge beyond DMS itself.

The trade-off is the one every agentic wrapper carries: it removes the friction that was doing quiet quality control. The assessment report used to be the thing you read before you committed to a target; when an agent generates and exports in one pass, the report becomes an artifact nobody opens. Keep it as a gate. Treat the converted DDL like a pull request from a fast junior: diff it against the source, run it against a scratch target, and route the stored procedures, functions, and triggers the agent "helped with" through a human who knows the application. The engine pairs where this bites hardest are the ones with the most dialect in them, Oracle or SQL Server to PostgreSQL, which are exactly the pairs most people are converting.

AWS Transform expert mode: fewer questions, same approvals

The migration job flow in AWS Transform was built around a chat that asks you things in order. Expert mode, added the same day, lets you supply those answers up front so the agent completes the migration steps with minimal interaction. For a team running its tenth wave this is the right default; the questions were the same every time and the answers lived in a spreadsheet anyway.

What does not change is the shape of the risk. The agent still needs source servers and network resources tagged CreatedBy: AWSTransform and ATWorkspace: <workspace_id> before it will operate on them, still allows one inventory import per target account and Region at a time, and still routes deployments through the Approvals tab, where only Admin-role users can confirm. Stopping a running job and restarting it begins from scratch. So expert mode is not autopilot; it is a shorter conversation in front of the same gates. Use it once your inventory file is stable, and keep the validate-and-confirm inventory step as the point where a human reads the instance types and subnets one more time.

Security Hub starts knocking from outside

Network Scanning, launched 8 July, probes your resources from the public internet to confirm actual reachability. It finds public IPs, virtual machines, and load balancers across AWS and Azure, identifies the reachable ports and the services behind them, and emits one finding per reachable port, with evidence. Security Hub Exposures then correlates those with configuration findings. It is included with Security Hub Essentials at no extra cost, on by default for new customers, and existing customers switch it on per account and Region or across the organization with a configuration policy.

This is a different question from the one Security Hub was already answering. "Could this be reachable, given the security groups and route tables" is a model of your network; "is this reachable, I just connected to it" is a measurement. The two disagree more often than teams expect, in both directions: a permissive security group behind an NLB that only forwards 443, or a reachable port nobody modelled because it arrived via a peering you forgot. The failure mode is volume. If you run intentional public ALBs on 443 you will get a finding for each one, and the temptation will be to suppress the whole control. Do not. Suppress by resource, keep the control, and let Exposures do the correlation that tells you which of the open ports are also on a host with an exploitable CVE.

The GPU fee cut, and what it says about Auto Mode

EKS Auto Mode and ECS Managed Instances both cut their management fees on accelerated instances: 35% on G-series, 60% on P-series and AWS Trainium, effective 1 July and applied automatically. EKS Auto Mode also picked up parallel image pulling on GPU instances with local NVMe and accelerator-aware node repair; ECS Managed Instances surfaces GPU utilisation, memory, and temperature in Container Insights and replaces instances on GPU hardware failure.

The management fee was the honest tax on letting AWS run your data plane, and on P-series it was large enough that cost-conscious teams kept GPU node groups self-managed while putting everything else in Auto Mode. Cutting it 60% narrows that gap to where the operational argument wins for most teams. Check the per-instance rate table on the pricing page: the percentages are per family and your bill is per instance hour.

Also this week

  • Aurora DSQL change data capture is GA, streaming inserts, updates, and deletes to Kinesis Data Streams, with Firehose onward to S3, Redshift, or OpenSearch; the announcement is silent on ordering and delivery semantics, so build idempotent consumers until the docs say otherwise.
  • OAuth for the AWS MCP Server: agents sign in through the browser with the same identity you use for the console, get short-lived tokens, and a headless flow uses create-oauth2-token-with-iam; new IAM condition keys and CloudTrail fields come with it.
  • Claude apps gateway for AWS, a self-hosted control plane that puts OIDC sign-in, managed settings, and per-user spend caps in front of Claude Code and Claude Desktop, routing inference to Bedrock. More on the agent side at ercan.ai.
  • Loom for AWS, open source in AWS Labs: a management UI and API for deploying Strands agents to AgentCore Runtime with RBAC, ABAC, and enforced tagging.
  • SageMaker Studio and Hugging Face integration: one-click deploy or customise, with default G5, G6, and G4dn quota for verified customers.
  • S3 Vectors reached both AWS GovCloud (US) Regions.
  • EC2 Region expansion: C8ine in Frankfurt and I7ie in Hyderabad.
  • Sandbox Environments on AWS Builder Center: free pre-provisioned accounts for workshop exercises, eight hours each, one per week, as the platform turned one.

The throughline

Three of this week's launches hand a formerly interactive workflow to a machine: schema conversion to an IDE agent, migration waves to expert mode, agent sign-in to OAuth. Each removes a step where a human used to pause and look. The announcements do not remove the checkpoints, the assessment report, the inventory confirmation, the Approvals tab, they just make it possible to rush past them. Network Scanning is the counterweight: a measurement, not a model, of what the rest of the week's automation left exposed. If you adopt the first three, adopt the fourth.

Read this next

For the AI and agent reading of the same weeks, the field notes live at ercan.ai. Consulting on AWS, cloud, CI/CD pipelines, Linux, GitHub/GitLab, and more runs through ercanermis.com.

References