Off-by-none: Issue #362

April 21, 2026

Mo’ Models, Mo’ Problems ⚠️

In our previous issue, AI started breaking things faster than we can defend them, AWS launched an agent registry, and S3 kinda became a filesystem. This week, Claude gets a major upgrade, AWS makes AI costs more visible, and Cloudflare goes all-in on agents. Plus, we've got some amazing cloud, serverless, and AI content from the community.

News & Announcements

Anthropic announced Claude Opus 4.7 this past week as their latest push towards world domination. Early signals point to serious gains in software engineering, especially for long-running tasks, plus stronger vision support. AWS wasted no time rolling it out in Bedrock.

AWS also introduced granular cost attribution for Amazon Bedrock, which is a big step toward actually understanding AI spend. Cost control and observability for LLMs is still pretty messy, and being able to map usage down to IAM users and roles starts to make that problem a lot more tractable.

Amazon Aurora Serverless is getting up to 30% better performance with smarter scaling, while still keeping the scale-to-zero promise. There’s a deeper dive from the team here if you want the details. I like this direction.

AWS also announced general availability of AWS Interconnect, kicking things off with Google Cloud. Dedicated bandwidth between clouds is becoming a thing, with Azure and Oracle Cloud Infrastructure expected to follow later this year. Let the homogeneity begin.

Anthropic introduced routines in Claude Code, which basically turns repeatable development workflows into something you can automate. Feels like another positive step toward making agents more useful in day-to-day dev work. They also highlighted what people are building in their ecosystem with their latest hackathon winners. No fluff, they're all practical AI solutions that address real pain points. 🤷

It was Agents Week over at Cloudflare last week, and they shipped a lot. The full rundown of launches is here, but there were a few standouts: AI Search as a core primitive for agents, Flagship to bring feature flags into the agent era, Agent Memory, and a new email service for agents in public beta.

Not on my 2026 Bingo card, but Apple announced that Tim Cook is stepping into the Executive Chairman role at Apple, with John Ternus taking over as CEO. Big shift for one of the most stable leadership runs in tech. I'm sure it has nothing to do with Apple Intelligence. 😬

And in case you missed it, the recent Vercel hack highlights a growing pattern in cybersecurity. Third-party AI tooling accessing internal systems is introducing a whole new threat model. One that most teams aren’t even aware of, never mind prepared for.

If your incident response still involves five tabs, three tools, and someone asking “who’s on point?”, it might be time to rethink things. incident.io is an all-in-one platform that runs Slack and Teams native, so you can declare, manage, and resolve incidents without leaving the conversation. It handles the busywork too, auto-assigning roles, kicking off workflows, and even surfacing insights from past incidents so you don’t keep fixing the same problem twice. Definitely worth a deeper look if you want faster response times without adding more process: incident.io. Sponsored

Tutorials

• Lambda Managed Instances: A Working Demo and the Math Behind It by Eric Johnson
• Transform retail with AWS generative AI services by Bhavya Chugh
• The Hidden Cost of AWS Lambda SnapStart for Python, and How I Fixed It with Durable Functions by Jaya Ganesh
• Best practices for using Claude Opus 4.7 with Claude Code
• Power video semantic search with Amazon Nova Multimodal Embeddings by Amit Kalawat
• Serverless applications on AWS with Lambda using Java 25, API Gateway and DynamoDB - Part 6 Using GraalVM Native Image by Vadym Kazulkin
• Accelerate database migration to Amazon Aurora DSQL with Kiro and Amazon Bedrock AgentCore by Noorul Mahajabeen Mustafa
• Using AWS Lambda Extensions to Run Post-Response Telemetry Flush by Melvin Philips
• Serverless applications on AWS with Lambda using Java 25, API Gateway and Aurora DSQL - Part 5 SnapStart + full priming by Vadym Kazulkin

Reads

Navigating the generative AI journey: The Path-to-Value framework from AWS
AWS tries to put some structure around the chaos with a “Path-to-Value” framework. It’s less of a step-by-step guide and more of a reminder that AI adoption is messy, multidimensional, and mostly about tradeoffs between value, risk, and organizational reality.

Moving past bots vs. humans
The bot vs human model is breaking down fast. Cloudflare is leaning into intent over identity, which feels like the right direction as agents start acting more like users and users start looking more like bots.

The AI engineering stack we built internally — on the platform we ship
Always interesting when a company dogfoods its own stack at scale. Cloudflare’s setup is a good look at what a modern AI platform actually needs when you’re pushing billions of tokens and not just running demos.

Multi-Agent AI in Production | Taskade Engineering (2026)
Three years into multi-agent systems and the same problems keep showing up. Memory, coordination, and agents getting stuck in loops. Good practical patterns here, especially if you’ve already hit these walls.

Why AWS Certified GenAI Developer stands apart from other AWS certs
Anwaar Hussain points out that this cert is less about knowing AI and more about wiring it into real systems. Which is probably the right shift, because building with AI is quickly becoming more of an architecture problem than a modeling one.

Lessons I learned building a memory-aware agent with Amazon Bedrock AgentCore Runtime
Memory is still the hardest part of agent design. Amit Kayal gives us a solid walkthrough of scoping, lifecycle, and not blowing up your prompts while trying to make agents feel stateful.

Learnings from conducting ~1,000 interviews at Amazon
Steve Huynh shares a good reminder that hiring is its own system with its own signals. If you don’t understand what a company actually optimizes for, you’re probably optimizing for the wrong thing.

Podcasts, Videos, and more

Serverless Apache Airflow | Serverless Office Hours
Airflow, but make it serverless. John Jackson and Kamen Sharlandjiev breakdown when MWAA actually makes sense versus just reaching for Step Functions, especially once you factor in cost, scaling, and how much orchestration complexity you really need.

Building, Managing & Governing APIs on AWS
APIs aren’t just for humans anymore. Giedrius Praspaliauskas covers the full lifecycle on AWS, but the interesting part is how API strategies are evolving to support agents, not just apps. Same primitives, very different consumers.

New from AWS

• MSK Replicator now supports replication from external Apache Kafka clusters to MSK Express Brokers
• Amazon MSK Replicator now supports enhanced consumer offset synchronization for bidirectional replication
• Amazon MSK Replicator now supports log forwarding for replication visibility
• Amazon DocumentDB (with MongoDB compatibility) now supports in-place upgrade from version 5.0 to 8.0
• Amazon ECR Pull Through Cache Now Supports Referrer Discovery and Sync
• Amazon CloudWatch now supports cross-region telemetry auditing and enablement rules

Developer Tools

brognilucas/sls-testing by Lucas Brogni
Typed, composable testing utilities for AWS Lambda from Lucas Brogni that provides event builders and Jest matchers for Lambda functions.

pujaaan/simple-cdk by pujaaan
A thin runtime over CDK that scans your folders, runs adapters in a deterministic three-phase pipeline (discover → register → wire), and emits real CDK constructs.

ToolSimulator: scalable tool testing for AI agents by Darren Wang
An LLM-powered tool simulation framework within Strands Evals to thoroughly and safely test AI agents that rely on external tools, at scale.

Final Thoughts 🤔

It’s getting a lot easier to build powerful systems.

Models are getting better at real work. Agents are starting to handle meaningful workflows. And the infrastructure around all of this is finally catching up, from cost visibility to orchestration to deployment patterns.

But the gaps are still there.

We’re wiring these capabilities into systems that were never designed for autonomous behavior. Giving tools access to internal systems. Letting agents make decisions across boundaries that used to be tightly controlled. And in some cases, we’re doing it faster than we understand the implications.

That’s where things start to break.

The Vercel incident isn’t an outlier. It’s a preview. A glimpse into what happens when powerful models meet loosely defined boundaries and third-party integrations. The tooling is evolving quickly, but the assumptions behind our systems haven’t fully caught up yet.

At the same time, you can see the industry starting to respond.

Better cost attribution. More structured agent workflows. Dedicated primitives for memory, search, and control. Even multicloud connectivity is starting to blur the lines between platforms. It’s not just about building faster anymore, it’s about building systems that can actually support what we’re asking them to do.

Still early. Still messy. But the pattern is emerging. More power, more abstraction, and more responsibility to get the boundaries right.

Because “Mo’ Models, Mo’ Problems” isn’t really a joke. It’s just the beginning.

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.