Off-by-none: Issue #359

March 31, 2026

So Long App Runner 🪦

In our previous issue, AWS gave agents real-time streaming and shell access, MCP was thought to be dead, and the distance between idea and deployment continued to shrink while quality became optional. This week, App Runner heads to maintenance mode, agents start looking like real software, and efficiency might become the new AI race. Plus, there's lots of awesome cloud, serverless, and AI content from the community.

News & Announcements

AWS finally did it. 😢 App Runner is heading into maintenance mode on April 30, 2026. It had its problems, but the free load balancer was hard to beat. Now it looks like the future belongs to Amazon ECS Express Mode (which requires a separate load balancer). If you’ve been using it as a trusted platform layer, and chances are you weren't, now might be a good time to reassess.

Despite the latest deprecations, AWS continues to compress the distance between idea and execution. You can now spin up and connect to an Amazon Aurora PostgreSQL database in seconds, which feels less like infrastructure provisioning and more like an on-demand service. Combine that with Lambda Managed Instances scaling up to 32 GB of memory and 16 vCPUs plus a new file descriptor limit of 4,096, and LMIs start looking like more than just a graduation path.

Step Functions also picked up 28 new service integrations, including Amazon Bedrock AgentCore, a clear signal that agent orchestration is becoming a first-class citizen. That pairs nicely with AgentCore’s other evolutions: managed session storage for persistent agent filesystem state and general availability of AgentCore Evaluations. Agents aren’t just stateless prompt loops anymore. They’re getting memory, persistence, and scoring systems that start to look like actual software systems.

On the model side, Palmyra Vision 7B is now available on Amazon Bedrock, adding more multimodal options to the growing Bedrock ecosystem. And if you’re building voice interfaces (like I am right now), Amazon Polly’s new bidirectional streaming brings real-time conversational synthesis into the mix. Adding voice interactions will change your life.

Outside of AWS, efficiency is the new AI battleground. Google’s TurboQuant research pushes model compression to extremes, hinting that smaller, cheaper models running locally might end up being more important than the cloud-based ones. Meanwhile, developer workflows keep shifting as tools like Auto mode for Claude Code lean further into autonomous execution, and enterprises get more guardrails with the Claude Compliance API. These seem like small steps, but they are filling lots of gaps.

And I can't believe I missed this a few weeks ago! Dash0 is acquiring Lumigo, making a bet that “agentic observability” is going to be its own category. You can’t run fleets of semi-autonomous systems without knowing what they’re doing, and traditional logs and traces aren’t going to cut it.

In case you missed it, and given the download numbers, you probably did, Sora is shutting down. It's a good reminder that as fake and curated as social media may be, there is a limit to the amount of AI slop users are willing to take.

Finally, a bit of the end of an era as the FullStack Bulletin bids farewell. Newsletters come and go, but the good ones reshape how we think about this space.

Tutorials

• Upgrading AWS CLI From v1 to v2 Using the Migration Tool by Ahmed Moustafa
• Building age-responsive, context-aware AI with Amazon Bedrock Guardrails by Pradip Kumar Pandey
• Accelerating custom entity recognition with Claude tool use in Amazon Bedrock by Kimo El Mehri
• AWS Lambda Now Knows Where It Lives: Explore New AZ Metadata by Darryl Ruggles
• Serverless applications on AWS with Lambda using Java 25, API Gateway and DynamoDB - Part 3 Introducing Lambda SnapStart by Vadym Kazulkin
• Serverless applications on AWS with Lambda using Java 25, API Gateway and Aurora DSQL - Part 3 Introducing Lambda SnapStart by Vadym Kazulkin
• The 15-Millisecond AI: Building "Pre-Cognitive" Edge Caching on AWS by Dhananjay Lakkawar
• Real-Time Streaming from Claude Code on AgentCore to the Browser Using AppSync Event API by Kota
• Serverless applications on AWS with Lambda using Java 25, API Gateway and Aurora DSQL - Part 2 Initial performance measurements by Vadym Kazulkin
• Multi-Agent Systems on AWS Lambda with Durable Functions by Gunnar Grosch
• Prevent Sensitive Data Leaks in Amazon CloudWatch Logs by Yaron Ben Ezra

Reads

Architecting for agentic AI development on AWS
Alan Oberto Jimenez covers architectural patterns for AI agents in development workflows. They walk through system architecture approaches (local emulation, hybrid testing, preview environments) and codebase patterns (domain-driven design, layered testing) that enable agents to autonomously write, test, and deploy code with rapid validation cycles.

Agents don't know what good looks like. And that's exactly the problem.
Luca Mezzalira reflects on a recent Neal Ford and Sam Newman fireside chat about AI agents. This is an extremely well-written and enlightening post that deserves your attention.

Your AI agents are a security nightmare
Great post by Allen Helton on treating AI agents as first-class principals with proper identity management. His approach using Teleport for session-bound AWS credentials makes a lot of sense, especially the parallel he draws to how we botched security with early Lambda deployments.

The Heirloom Syntax: Why AI Monocultures Threaten the Future of Innovation
Great post from Brian Tarbox on preserving "systemic diversity" in an age of AI commoditization. I totally agree that our value as technologists isn't in content volume anymore (machines won that game), but in maintaining the unique voice and insights that come from actual experience in the trenches.

How we use Abstract Syntax Trees (ASTs) to turn Workflows code into visual diagrams
André Venceslau explains how Cloudflare built a diagram system that parses workflow code using ASTs. Great deep dive if you want to understand the technical challenge of statically analyzing minified JavaScript to track Promise relationships and parallel execution, then rendering those into visual diagrams that show execution order and branching. Cool stuff.

Why I Collapsed 50+ AWS Lambdas Into a Single API
Harshit walks through consolidating 50+ Lambda functions into a single API to address cold start multiplication and VPC ENI exhaustion. I've built A LOT of serverless APIs over the years, and I'm a big proponent of this approach. There are still good reasons for single-purpose functions, but a Lambdalith using Hono for internal routing is extremely easy to reason about.

AI is creating the first generation of cognitively outsourced humans
The cognitive offloading problem we're facing with generative AI is definitely real. Confusing fluent AI output with actual understanding is where most people get fooled (and even worse if they believe its sycophancy), and I totally agree that we need to treat AI as a tool to sharpen judgment, not to replace it.

Podcasts, Videos, and more

AWS MCP server | Serverless Office Hours
Praneeta Prakash and Claudiu Popa join Julian Wood to explore the AWS MCP Server. It combines natural language interfaces with AWS best practices through Agent SOPs. You'll see practical examples of provisioning infrastructure, troubleshooting, and building Lambda durable functions while maintaining IAM controls throughout.

New from AWS

• Amazon Athena launches Capacity Reservations in additional regions
• Amazon CloudWatch now supports multi-account and region log centralization based on data source
• Access Cluster Insights through the Amazon OpenSearch Service Console and Amazon EventBridge events
• Amazon CloudWatch Logs now supports data protection, OpenSearch PPL and OpenSearch SQL for the Infrequent Access ingestion class
• Amazon Timestream for InfluxDB Now Supports Advanced Metrics
• The AWS Advanced JDBC Wrapper now supports automatic query caching with Valkey
• Aurora DSQL launches connector that simplifies building Ruby applications
• Amazon Managed Service for Apache Flink now supports Apache Flink 2.2
• AWS Transform custom announces general availability of automated codebase analysis
• AWS Transform custom introduces new AWS-managed transformations to modernize code at scale
• Amazon Bedrock AgentCore adds support for Chrome policies and custom root CA
• AWS Security Agent on-demand penetration testing is now generally available
• AWS DevOps Agent is now generally available
• Amazon CloudWatch Logs introduces lookup query command
• Aurora DSQL launches new connectors that simplify building .NET and Rust applications

Developer Tools

gunnargrosch/durable-viz
Durable-viz parses AWS Lambda Durable Functions code to generate Mermaid flowcharts without requiring deployment. The tool supports TypeScript, JavaScript, Python, and Java, detecting all SDK primitives like step, invoke, parallel, map, and wait. It's available as an NPM package or VS Code extension with features like click-to-navigate and PNG export.

How I Search 10,000+ AWS GitHub Repos in 10 Seconds by Ajit
Ajit built a search engine that indexes over 10,000 AWS GitHub repositories using hybrid search (BM25 for exact matches, FAISS for semantic understanding). The system auto-indexes twice daily via EventBridge and uses Bedrock to classify repos across 22 metadata dimensions.

Introducing Loom, an agent platform by Heeki Park
Heeki Park's Loom is an opinionated agent platform built on AWS that integrates Strands Agents SDK, Bedrock, and AgentCore with enterprise guardrails. The platform uses configuration-driven agent creation with mandatory tagging and RBAC/ABAC controls, avoiding runtime code generation for security.

Final Thoughts 🤔

It feels like infrastructure is starting to shift in a more fundamental way. Not just bolting AI onto existing systems, but weaving it into everything: orchestration, storage, compute, observability. Agents with memory, evaluation loops, persistent state. Voice interfaces that feel real-time. Even the way we think about databases and execution environments is starting to change.

But we might be getting ahead of ourselves.

Luca’s point that agents don’t actually know what “good” looks like, and Brian’s argument for preserving systemic diversity instead of letting AI push everything toward a monoculture, are worth paying attention to. If every system is built the same way, trained on the same data, and optimized for the same outputs, we’re not really innovating, we’re just converging.

The stack is evolving fast. The question is whether we’re shaping it, or just letting it shape us.

See you next week,
Jeremy

I hope you enjoyed this newsletter. We're always looking for ideas and feedback to make it better and more inclusive, so please feel free to reach out to me via Bluesky, LinkedIn, X, or email.