Off-by-none: Issue #358

March 24, 2026

The Cost of Going Faster 🚄

In our previous issue, S3 turned 20, AWS made agents more stateful and observable, and Claude expanded to a 1M token context window. This week, AWS gives agents real-time streaming and shell access, MCP might be dead, and the distance between idea and deployment shrinks while quality becomes optional. Plus, we've got lots of cloud, serverless, and AI content from the community.

News & Announcements

Amazon Bedrock AgentCore got an upgrade to make agents feel less like request/response loops and more like real systems. The runtime now supports WebRTC for real-time bidirectional streaming, which opens the door for low-latency, continuous interactions instead of polling APIs like it’s 2015. They also added shell command execution, which is both incredibly powerful and a little terrifying. Letting agents execute commands directly is how you get real work done… and also how you accidentally run a destructive Terraform command. 😬

On the MCP side, the AWS MCP Server (Preview) got enhanced monitoring and semantic search capabilities. Observability and retrieval are quickly becoming first-class concerns for agent systems, and this is AWS acknowledging that you can’t operate these things without both.

A smaller but interesting update: AWS Lambda now supports Availability Zone metadata. It’s one of those features you don’t think about until you need it, especially for latency-sensitive or fault-aware workloads. Then suddenly it matters a lot.

AWS also introduced the Nova Forge SDK, a new way to customize Nova models for enterprise use. This feels like AWS continuing to push into the “bring your own intelligence, but we’ll help you shape it” layer. At the same time, they announced a $12.5M investment to defend the open source ecosystem from AI threats. As AI starts to consume and replicate open source faster than it can be sustained, this is a problem that’s only going to get more attention.

Over at Anthropic, Claude keeps creeping closer to being an actual operating system for work. Computer use and dispatch means Claude can now take actions directly on your machine. It's in preview and they admit it'll get things wrong, but they're definitely keeping other labs on their toes.

Speaking of OpenAI, they announced plans to acquire Astral, continuing the trend of model companies pulling more of the stack in-house. Meanwhile, Netlify (remember Jamstack?) is leaning into the opposite side of that trend, launching a new way to turn prompts into production-ready web projects. The gap between “idea” and “deployed app” keeps shrinking, but I fear the gap between “high quality” and “slop” is getting even bigger.

Cloudflare also dropped a big one with dynamic workers for sandboxing AI agents up to 100x faster. Safe execution environments are becoming table stakes for agent systems, and doing that faster at the edge is exactly where this needs to go.

And finally, AWS announced its first 2026 Heroes cohort. Always great to see the community getting recognized. There are some smart, well-deserved folks on that list.

It’s been well over a decade since the first serverless functions, yet delivering near-native performance without cold starts or platform sprawl is still harder than it should be. Serverless on Akamai combines Akamai Functions and EdgeWorkers with a WebAssembly-based runtime that launches in microseconds and scales instantly across a distributed cloud-to-edge platform, letting you build APIs, microservices, and event-driven workflows that support AI inference and modern workloads. If you’re exploring architectures that keep latency low while reserving centralized compute for heavier processing, this is a platform worth checking out. Sponsored

Tutorials

• Integrating Amazon Bedrock AgentCore with Slack by Salman Ahmed
• Use RAG for video generation using Amazon Bedrock and Amazon Nova Reel by Nick Biso
• Overcoming LLM hallucinations in regulated industries: Artificial Genius’s deterministic models on Amazon Nova by Paul Burchard and Igor Halperin
• Build an AI-Powered A/B testing engine using Amazon Bedrock by Vijit Vashishtha
• Evaluating AI agents for production: A practical guide to Strands Evals by Ishan Singh
• Serverless applications on AWS using Lambda with Java 25, API Gateway and DynamoDB - Part 2 Initial performance measurements by Vadym Kazulkin
• Deploying an Image Recognition Service to AWS Lambda by Isaac FEI
• A Simple Like System for a Static Blog by Isaac FEI
• Extract data from Amazon Aurora MySQL to Amazon S3 Tables in Apache Iceberg format by Kunal Ghosh, Arghya Banerjee, Indranil Banerjee, and Vipan Kumar
• Anatomy of the .claude/ folder - A complete guide to CLAUDE.md, custom commands, skills, agents, and permissions, and how to set them up properly by Akshay Pachaar
• Handlebars with AWS Lambda Layers by Pathum Jayaweera
• AWS Lambda Durable Functions: Building Long-Running Workflows in Code by Gunnar Grosch

Reads

Production Is Where the Rigor Goes
I totally agree with Charity's take on why engineers should engage with production beyond just bugs and incidents. We should be constantly checking production to understand how our systems actually behave, not waiting for something to break. Observability isn't a nice-to-have, it's how we understand what we've built.

Strands vs. Claude Agent SDK: Two Very Different Bets on What “Agent” Means
Luca Bianchi's technical comparison of AWS Strands Agents and Claude Agent SDK covers the key architectural differences: Strands as orchestrator requiring custom tool implementations versus Claude SDK as full runtime with pre-built tools. Includes practical details on TypeScript support, multi-model usage, and security considerations for production deployments.

Introducing V-RAG: revolutionizing AI-powered video production with Retrieval Augmented Generation
Nick Biso introduces Video Retrieval-Augmented Generation (V-RAG), which combines image retrieval with video generation models to improve accuracy and reduce hallucination. The approach builds on existing AWS services and doesn't require model fine-tuning.

20 years in the AWS Cloud – how time flies!
Channy Yun walks through two decades of AWS evolution, from the early days of S3 and EC2 to modern serverless offerings like Lambda, Aurora Serverless v2, and Fargate.

State of the product job market in early 2026
Lenny's biannual tech job market report reveals some surprising trends: AI positions are (obviously) accelerating rapidly, but design roles have plateaued, and the Bay Area now claims over 23% of all PM openings. Worth reading if you're tracking talent flows in cloud-native companies.

AI Is Eating Open Source From the Inside
Really interesting read on how AI coding tools are fundamentally changing our relationship with open source. AI-generated contributions have 1.7x more issues, and I think the point about developers never visiting docs or filing real bugs anymore hits hard. We're outsourcing engagement to tools that don't actually participate in the ecosystem.

MCP Won. MCP Might Also Be Dead.
As someone who has faced the pain of both using (and building) bloated MCP servers, I totally agree that building protocol-agnostic tools that export to everything is the right approach. Write your logic once, export to MCP for ecosystem reach, export to native formats for teams wanting lower overhead. MCP isn't going away, but it's still worth avoiding betting everything on one standard.

Claude Code Best Practices: Lessons From Real Projects
Great post by Ran Isenberg on using Claude Code for real projects. His point that domain expertise matters more than the tool is spot on, especially his example of BMAD generating 36 user flows he didn't know he needed because he was asking the right questions about his domain.

Product management on the AI exponential
Cat Wu describes Anthropic's product management workflow for building with Claude: using Claude.ai for ideation, Claude Code for prototyping, and Cowork for knowledge work. The article breaks down principles like planning in short sprints and favoring demos over documentation as models improve faster than traditional product cycles.

Are AI agents actually slowing us down?
Great post by Gergely Orosz on the hidden costs of AI coding agents. The data from Amazon, Meta, and Uber showing increased outages and bugs should scare everyone. Productivity metrics don't tell the whole story when quality suffers.

The Convergence Problem: Rethinking the 2028 Global Intelligence Crisis
The real value of SaaS isn't just the software, it's the centralized responsibility for compliance, security, and integration. I'm increasingly convinced that AI won't eliminate this need, it'll just redistribute the complexity in ways we haven't fully mapped yet. This post articulates some of my thoughts.

Podcasts, Videos, and more

Analytics for Modern Data Lakes & AI | Serverless Office Hours
Scott Rigney, Vijay Jain, Theo Tolv, and Andy Warfield join Eric Johnson to show how Amazon Athena powers over one billion queries per week with sub-second starts and no infrastructure to manage.

Your .NET Worker Service Is Slower and More Expensive Than AWS Lambda!
James Eastham walks through migrating .NET worker services from traditional polling to Lambda-based SQS processing. Includes practical code examples, failure handling patterns, batch processing configurations, and a load simulation showing how Lambda scales compared to traditional workers.

New from AWS

• Amazon Polly expands Generative TTS engine with 10 new voices, 2 new regions, and Bidirectional Streaming API
• AWS DataSync now supports AWS Secrets Manager for all location types
• NVIDIA Nemotron 3 Super now available on Amazon Bedrock
• AWS adds support for NIXL with EFA to accelerate LLM inference at scale
• AWS Security Agent now supports downloading penetration testing reports
• Minimax M2.5 and GLM 5 models now available on Amazon Bedrock
• Amazon SageMaker Unified Studio adds custom metadata filters
• AWS Config launches 75 new managed rules
• Simplified permissions for Amazon S3 Tables and Iceberg materialized views
• Amazon Corretto 26 is now generally available

Upcoming Events

March 26, 2026 - AI Codecon: Software Craftsmanship in the Age of AI by O'Reilly Media, Inc.

Final Thoughts 🤔

There’s a pattern emerging here, and it’s getting hard to ignore.

Agents are no longer just calling APIs. They’re streaming, executing, observing, retrieving, and operating inside real environments. The shift from “model as a function” to “model as a system” is happening faster than most of us suspected.

But as the surface area expands, so does the blast radius.

Real-time streaming and shell access make agents more useful, but also more dangerous. Faster sandboxing helps, but it’s a response to a problem we’re actively accelerating. Observability and retrieval are becoming first-class not because they’re nice features, but because without them, these systems are unmanageable.

At the same time, the distance between idea and deployment keeps collapsing. Prompts to production is no longer the exception, it's the norm. The constraint is no longer building. It’s judgment.

And that’s the tension.

We’re building systems that can do more than ever, faster than ever. But the signals from production, open source, and even large engineering orgs are clear. Speed without rigor creates fragility. Automation without accountability creates noise.

The next phase isn’t just about making agents more capable. That part is already happening.

It’s about making them trustworthy.

And that’s a much harder problem.

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.