Off-by-none: Issue #356

March 10, 2026

OpenClaw, Now with More Cloud! 🦞☁️

In our previous issue, DSQL kept checking boxes, serverless PHP got a major upgrade, and Dario Amodei tried blocking production of the T-1000. This week, AWS helps your AI agents go rogue, Anthropic puts the AI code review market on notice, and GPT-5.4 arrives just in time for that Department of War contract. Plus, we've got a ton of awesome contributions from the community.

News & Announcements

Amazon is now offering OpenClaw on Lightsail. Apparently AWS would like to contribute its servers and bandwidth to the noble cause of helping your personal AI agents go fully rogue. At least now you don't have to self-host the chaos or roll the whole stack yourself.

Policy in Amazon Bedrock AgentCore is now generally available, which intercepts agent-tool traffic and evaluates each request before allowing or denying access. It's a great feature, but it also means the “simple” agent architecture now comes with more moving parts to reason about.

AWS also announced a new Kiro power to help with durable functions development (Accelerate Lambda durable functions development with new Kiro power). One of the downsides of creating a new state management programming interface is that the LLMs aren't trained on it, so giving Kiro a dedicated power to understand it is probably the only way to keep developers from giving up on it.

The Guardian reported on drone strikes and the implications for Gulf-hosted AI infrastructure. It does make an uncomfortable kind of sense that data centers may soon need their own air defense systems, which is a sentence that should concern everyone. The only thing worse would be letting the AI run those defenses autonomously so it can protect itself from intentional human shutdowns. 🤖

Anthropic announced Code Review for Claude Code, and as is the norm with Anthropic, they just put every other AI-powered code reviewer on notice. They keep shipping features that feel less like demos and more like a direct threat to entire categories.

And finally, OpenAI rolled out GPT-5.4, just in time for that big Department of War contract, because nothing says progress like ever more capable AI landing exactly where the defense budget can appreciate it most.

Tutorials

• Serverless applications with Java and Aurora DSQL - Part 4 Understand DPUs in Explain Analyze Verbose by Vadym Kazulkin
• MCP ServerLESS: Building MCP Functions on AWS Lambda with TypeScript and AWS SAM by Kelwin Ferreira
• Automate AWS Lambda Runtime Upgrades with AWS Transform custom by Venugopalan Vasudevan and Gokul Sarangaraju
• Running Small AI Models on Lambda with ONNX by Ronny Roeller
• Common workflow patterns for AI agents by Anthropic
• Safe AWS Lambda Durable Functions with XState by szymon-szym
• Serverless Bedrock: How I invoke Claude from Lambda in warrantyAI by Harish Aravindan
• Building and Shipping Apps Faster with Kiro’s Agentic Power by Awedis Keofteian
• AWS Lambda Managed Instances with Java 25 and AWS SAM – Part 5 Lambda function initial performance measurements by Vadym Kazulkin
• A Serverless Recipe Assistant with Bedrock AgentCore, Knowledge Bases, and S3 Vectors by Darryl Ruggles
• OpenClaw on AWS Bedrock AgentCore: Secure and Serverless by Matias Kreder
• Circuit Breakers on AWS Lambda: Why In-Memory State Silently Fails by Gunnar Grosch
• AWS Lambda Managed Instances with Java 25 and AWS SAM – Part 6 Lambda function performance improvement approaches by Vadym Kazulkin
• Building custom model provider for Strands Agents with LLMs hosted on SageMaker AI endpoints by Dan Ferguson

Reads

The Complete Guide to Building Skills for Claude
This guide is a fantastic resource whether you’re building agents or assistive AI workflows, showing exactly how to package repeatable processes into reusable skills so your AI stops improvising every task and starts executing like a well-trained teammate. 🔥

How I Dropped Our Production Database and Now Pay 10% More for AWS
I'm not sure I agree with letting AI agents anywhere near production Terraform in the first place. The post describes running terraform apply through Claude Code without manually reviewing the plan, then being surprised when a long list of resources got destroyed. That feels like the automation was working exactly as designed (just on the wrong state file).

Your LLM Doesn't Write Correct Code. It Writes Plausible Code.
This is a great take on LLM-generated code from Hōrōshi, who demonstrates that the real problem isn't syntax errors but plausible incorrectness. The key insight: LLMs work best when you define acceptance criteria before generating code, not after.

AI Doesn’t Reduce Work—It Intensifies It
An eight-month study about AI productivity confirms exactly what I've been seeing: AI tools don't give us more free time, they just let us do more work in the same hours, and that's not sustainable long-term.

Custom AWS Serverless Blog Built in 10 Hours with AI
Darryl Ruggles built a production serverless blog in 10 hours using Claude Code. But while AI accelerates development, it still requires experienced judgment for architecture and security decisions, which feels like exactly the right framing for where we are with these tools. For now.

Trust Is the Architecture (Part 1)
Daniele Frasca explores governance and trust considerations for serverless at scale, covering multi-team coordination, event-driven architecture contracts, data residency compliance, and audit requirements. All the things that matter much more than the technology you choose.

The AI political resistance has arrived
What do you do when AI threatens to destroy humanity? Write a strongly worded letter! 🙂 The Pro-Human AI Declaration, signed by an unprecedented coalition of labor, religious, and political organizations, establishes five principles for human-centered AI development. And it was all cloak and dagger with a secret meeting that intentionally left out Big Tech reps. 🕵️

Improving skill-creator: Test, measure, and refine Agent Skills
Anthropic's new skill-creator features focus on testing workflows: parallel eval execution to avoid context bleed between runs, A/B comparison tools for iterating on improvements, and capability uplift metrics. AI making AI better, more predictable, and auditable.

Something big is happening in AI — and most people will be blindsided
Matt Shumer argues we're hitting a transformative AI moment where models can now handle complete jobs autonomously, not just assist with tasks. There's still a lot of nuance that goes into building software, but the scope and complexity of tasks AI can now handle is exponentially higher than a year ago.

Podcasts, Videos, and more

AWS Bites #153. LLM Inference with Bedrock
Really great deep dive from Eoin and Luciano on Bedrock's structured outputs and constrained decoding. If you're building anything that needs reliable JSON parsing, be sure to give it a watch.

Why .NET Devs Should Stop Writing Lambda YAML Forever
This tutorial from James Eastham demonstrates how CDK lets .NET developers define infrastructure using C# instead of YAML or CloudFormation templates. It covers building reusable constructs and automating the deployment of Lambda functions with proper permissions management.

New from AWS

• Amazon CloudWatch Logs announces increased query concurrency and API limits
• Amazon SageMaker Unified Studio launches support for remote connection from Kiro IDE
• Amazon SageMaker Unified Studio now supports AWS Glue 5.1 for data processing jobs
• Amazon OpenSearch Ingestion now supports Amazon Managed Service for Prometheus as a sink
• Amazon OpenSearch Ingestion now supports unified ingestion endpoint for OpenTelemetry data
• AWS SDK for .NET V3 Maintenance Mode Announcement
• AWS simplifies IAM role creation and setup in service workflows
• Mistral AI Devstral 2 now available on Amazon Bedrock
• Amazon EventBridge Scheduler now provides a higher default quota for the CreateSchedule API
• Amazon OpenSearch Service introduces capacity optimized blue/green deployments
• Amazon Redshift Serverless now maintains datashare permissions during restore
• Amazon Redshift introduces reusable templates for COPY operations
• Amazon Redshift introduces new array functions for semi-structured data processing
• AWS Elastic Beanstalk now offers AI-powered environment analysis

Thoughts from Social

Amazon is holding a mandatory meeting about AI breaking its systems. The official framing is "part of normal business." The briefing note describes a trend of incidents with "high blast radius" caused by "Gen-AI assisted changes" for which "best practices and safeguards are not… pic.twitter.com/XSXOSqALBN

— Lukasz Olejnik (@lukOlejnik) March 10, 2026

Interesting take on Amazon’s response to AI coding incidents. Adding human oversight and requiring senior sign-off sounds reasonable, but it also raises a bigger question: if a company with AWS-level engineering depth is still struggling to harden AI-assisted development workflows, what hope is there for the rest of us?

Developer Tools

OMDA - Organizational Memory Decay AI by Samyak Jain
AI-powered early-warning system that predicts and prevents critical knowledge loss in organizations.

Final Thoughts 🤔

I’ve been spending a lot of time lately continuing work on my local agent orchestrator, and it’s getting… scary good. 😬 Watching an army of agents coordinate work across isolated sandboxes and shared context without stepping on each other is both fascinating and slightly concerning.

I’m also still deep in the Context Engineering and AI memory management rabbit hole. The more time I spend working on this, the more convinced I am that proper memory architecture is going to be the difference between chaos and reliable AI.

And as always, my side projects are quietly making progress in the background. I’m pretty excited about one in particular. Hopefully I’ll have more to share soon.

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.