Off-by-none: Issue #355

March 3, 2026

Saving Us From Skynet 🤖

In our previous issue, DSQL kept inching toward "new default" status, Claude kept launching startup killers, and FinOps consultants started turning their advice into software. This week, DSQL keeps checking boxes, serverless PHP gets a major upgrade, and Dario Amodei blocks production of the T-1000. Plus, we've got some awesome articles from the serverless and cloud community.

News & Announcements

AWS continues to make really smart moves with Aurora DSQL. This week, they launched a playground for interactive DSQL exploration, added support for Tortoise, Flyway, and Prisma, and shipped new integrations for VS Code SQLTools and DBeaver. This is exactly how you turn it into the default: reduce friction, meet developers where they are, and integrate with the tools we already use. DSQL has been checking off a lot of boxes for me lately.

OpenSearch Serverless introduces collection groups to help optimize costs for multi-tenant workloads. This is a genuinely great feature for SaaS architectures that need isolation without exploding infrastructure spend. That said, I still don’t love the overall pricing model. It’s better, but not quite where I’d like to see it for broader adoption.

For those of you still writing Java like it’s 1999, this one’s for you. The AWS Lambda Durable Execution SDK for Java is now available in developer preview. Durable execution patterns in Java on Lambda might not be trendy, but they’re very real in large enterprises. If that’s your world, I'm sorry. But also, this is a meaningful step forward.

Amazon Bedrock had a big week too. Batch inference now supports the Converse API format, which is 🔥. Converging on a single format across real-time and batch inference is absolutely the right move. Fewer abstractions. Fewer translation layers. Less messy glue code. Plus, server-side tool execution with AgentCore Gateway is now available, which is a major step toward fully-managed serverless agents. Let the platform handle the orchestration loop so we don't have to keep upgrading our agent frameworks every 15 minutes.

OpenAI and Amazon announced a strategic partnership (obviously). The gravitational pull between model providers and cloud platforms was always going to win. Alongside that, Bedrock launched an OpenAI-compatible Projects API, making migration and interoperability much simpler. Less friction, fewer rewrites, and one more signal that the model layer is becoming commoditized infrastructure.

On a more serious note, Anthropic published a statement from Dario Amodei regarding discussions with the Department of War. I agree with the principles outlined there. As AI becomes more capable, governance and alignment discussions are not optional. I feel like watching The Terminator, The Matrix, and WarGames needs to become a standard part of the War Department's new employee onboarding.

And to close things out, two awesome announcements. Valkey Admin brings visual cluster management to Valkey, which should make operating distributed cache clusters far less painful. And Bref 3.0 is officially out, continuing to make PHP on AWS Lambda a very real (and increasingly polished) option.

It’s been well over a decade since the first serverless functions, yet achieving near-native performance without platform sprawl or cold-start penalties is still harder than it should be. Akamai Functions changes the equation with a WebAssembly-native runtime built on open standards like Spin and WASI, offering microsecond startup times and instant scale across a unified cloud-to-edge platform. Whether you’re writing in Rust, Go, Python, or TypeScript, you get a memory-safe, portable execution layer that integrates compute (CPU/GPU), storage, and security before requests ever hit your core infrastructure. If you’re rethinking how to build high-performance, event-driven systems without the operational overhead of traditional serverless, this is definitely a platform worth exploring: Akamai Functions. Sponsored

Tutorials

• Understanding Lambda Tenant Isolation by Pubudu Jayawardana
• I Rewrote My Step Function as a Durable Function by Danielle Heberling
• How to Monitor Event Delivery in Amazon EventBridge by Renaldi Gondosubroto
• Streaming Bedrock Responses Through API Gateway and Lambda by Gunnar Grosch
• I Enabled Aurora Data API and My AI Agent Started Querying the Database Directly by Kohei Aoki
• AWS Lambda Durable Functions on Hexagonal Architecture: The Pattern You’ve Been Looking For by Matheus das Mercês
• Serverless applications with Java and Aurora DSQL - Part 5 Support for identity columns and sequence objects added by Vadym Kazulkin
• AWS Lambda Managed Instances with Java 25 and AWS SAM – Part 4 Monitoring, unsupported features, challenges and pricing by Vadym Kazulkin
• 9 Tips for Reducing API Latency in Agentic AI Systems by J. Simpson
• Architecting Low-Latency, Real-Time AI Voice Agents: Challenges & Solutions

Reads

We deserve a better streams API for JavaScript
Great post by James Snell on rethinking JavaScript streams. His async iterable approach shows 2x to 120x performance improvements across all major runtimes, and the pull-based model makes so much more sense than the current Web streams design (no hidden buffering, no eager evaluation). Given that these issues stem from design decisions made before async iteration existed in JavaScript, it's probably time to rethink the design.

The serverless lie: Why I refuse to default to next js
While I certainly agree with Jeremy's take on Next.js and met-frameworks, I'm not convinced that the default should be FastAPI backends hosted on a VPS. Serverless still has sharp edges, but the ecosystem goes well beyond just the function execution focused on in this article. I'm serverless-first until the architecture requires something different.

Claude Built My Wix Website in 3 Hours - Is SaaS Dead?
Ran Isenberg experimented with using Claude to replace his Wix website, successfully migrating to a custom AWS setup (CloudFront/S3) in 3 hours. He walks us through his AI-SDLC approach and discusses what worked, what didn't, and where the real challenges lie in long-term maintenance.

From Orchestration to Authority
Really interesting take on why traditional workflow orchestration falls apart when you introduce AI agents. The key insight here is that autonomy isn't about smarter agents, it's about explicitly defining authority boundaries and governance. This is the drum I've been beating for quite some time now.

Every service I build will die
Great post by Ollie Church on building services that are designed to be replaced. The insight that flexibility comes from architecture (small Lambda functions, EventBridge routing) rather than upfront planning is spot on. This is serverless forcing us to think differently, just like it was intended to (maybe I'm rewriting history a bit, but it fits the narrative 😉).

6,000 AWS accounts, three people, one platform: Lessons learned
Ben Freiberg breaks down how his team manages 6,000 tenant accounts with only a three-person team. Great breakdown of their investment in platform engineering and how they built internal tools to abstract deployment complexity away. Same idea we took with Ampt to maintain complete account-level isolation for security and cost controls.

Code Review Is Not About Catching Bugs
David Poll argues that code review's purpose is validating design intent and architectural coherence, not bug detection. I tend to agree, but I think it depends heavily on what your process is. If you're just skimming code in GitHub, it's unlikely that any developer is going to catch something unless it's glaringly wrong. If you're checking out PRs and testing them as part of your code review, well that's a different story.

Babe Wake Up, AI Killed My Job. Again?
Vishwas examines whether AI is truly replacing software engineers, arguing that code generation is the easy part while problem understanding and architectural decisions remain human responsibilities. I've said it before, and I'll say it again: most engineers know this, most CEOs don't.

Podcasts, Videos, and more

Immutable Patterns of System Design; From Monolith to Agentic AI - James Eastham - NDC London 2026
James Eastham's NDC London talk walks through system design patterns that remain relevant across different architectural paradigms. He discusses some really helpful reliability and performance patterns that work whether you're maintaining legacy systems or building with the latest technologies.

Mitchell Hashimoto’s new way of writing code
Mitchell Hashimoto discusses HashiCorp's journey from open-source projects to a durable business, plus his current approach to using AI in development. The conversation also covers his perspective on the major cloud providers and what's changing for engineers in this AI-native era.

AI Agents Will Do What Cloud Did to Ops - Are You Ready?
Adrian Cockcroft and Jamie Dobson explore the parallels between cloud's automation of ops roles and AI's potential impact on programming jobs. They also discuss organizational adoption patterns and the mindset needed for navigating technological change. Adrian predicted the cloud shift, so probably worth spending some time listening to this one.

New from AWS

• AWS Observability now available as a Kiro power
• AWS WAF announces AI activity dashboard for visibility into AI bot and agent traffic
• Amazon Lightsail expands blueprint selection with a new WordPress blueprint
• Amazon CloudWatch logs centralization rules now support customizable destination log group structure
• Amazon S3 now provides AWS source region information in server access logs
• Amazon Cognito enhances client secret management with secret rotation and custom secrets
• Amazon EventBridge Scheduler adds resource count metrics for quota monitoring

Thoughts from Social

Turn Dependabot off | Hacker News
Dependabot is great at keeping you busy and terrible at telling you whether you’re actually in danger. This thread nails the real upgrade path: stop treating version numbers like a threat model and start prioritizing reachability, call paths, and context, otherwise your biggest vulnerability is alert fatigue.

Final Thoughts 🤔

I spent some time experimenting with OpenClaw, then went heads down building my own local agent orchestrator so I could get extremely prescriptive about context management and security. Every agent runs inside an isolated Docker sandbox with limited tool and file access, plus explicit guardrails to keep them on a short leash. I’ve wired the whole thing through a local MCP server so agents can communicate with one another and safely fan out work, inheriting permissions and file access from the parent agent. It’s been fascinating seeing how far I can (safely) push this thing. I’m also getting very close to releasing a new side project I’ve been working on, so stay tuned for that. 🙂

Also, the Data API Client v2.1.4 is out. If you’re using Aurora Serverless v2, please take a look and send me your feedback. 🙏

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.