Off-by-none: Issue #361

April 14, 2026

S3 is Still Not A File System 🤷

In our previous issue, Claude leaked its secrets, AWS locked down S3, and Cloudflare made a run at WordPress. This week, AI starts breaking things faster than we can defend them, AWS launches an agent registry, and S3 becomes a filesystem (sort of). Plus, we've got plenty of awesome content from the cloud, serverless, and AI communities.

News & Announcements

Anthropic just dropped Claude Mythos Preview, a gated research release (available on Amazon Bedrock) that’s already raising eyebrows. Early reports suggest it’s extremely good at finding vulnerabilities in long-standing systems, which is both impressive and a little terrifying. They’re rolling it out slowly so people can prepare their security programs for AI-accelerated offense. The reality is AI is getting very, very good at breaking things, and we’re going to need to rethink how we defend systems that were never designed for this level of pressure.

AWS made a bunch of moves in Bedrock this week. They added cost allocation by IAM user and role, which is a big win for visibility into AI spend. They also introduced stateful MCP client capabilities in AgentCore Runtime and added OS-level actions to the AgentCore Browser. Good stuff.

AWS also launched the Agent Registry in preview. Centralized discovery and governance is going to be critical as agent sprawl becomes a thing (think shadow APIs on steroids). There’s a deeper dive on the vision in this post on managing agents at scale.

Big S3 news this week with the launch of S3 Files, which essentially puts a filesystem interface in front of your buckets. If you want more detail, the full breakdown is in this launch post. Still not a filesystem… but closer than ever before.

AWS Lambda response streaming is now available in all commercial regions, which is great to see. This feature has unlocked much better UX for real-time apps. I'm glad there's no longer a regional bottleneck.

If you still love PHP, you can now plug into Aurora DSQL with the new PHP connector. Not flashy, but extremely useful.

Anthropic also introduced Claude Managed Agents, which is basically a faster path to getting agents into production without wiring everything up yourself. Less plumbing, more doing.

Cloudflare announced Sandboxes GA. Giving agents their own isolated environments feels like table stakes at this point.

And finally, Vercel might be getting ready for Wall Street. Their CEO is signaling IPO readiness as AI agents drive growth. You can read more about it here.

Tutorials

• Embed a live AI browser agent in your React app with Amazon Bedrock AgentCore by Sundar Raghavan
• Manage AI costs with Amazon Bedrock Projects by Ba'Carri Johnson
• Seeing like an agent: how we design tools in Claude Code by Thariq Shihipar
• Serverless applications on AWS with Lambda using Java 25, API Gateway and Aurora DSQL - Part 4 SnapStart + DSQL request priming by Vadym Kazulkin
• Building intelligent audio search with Amazon Nova Embeddings: A deep dive into semantic audio understanding by Madhavi Evana
• Lambda Just Got a File System. I Put AI Agents on It. by Eric Johnson
• Serverless applications on AWS with Lambda using Java 25, API Gateway and DynamoDB - Part 5 SnapStart + full priming by Vadym Kazulkin

Reads

How I Use OpenClaw as My AI-Powered Personal Operating System
Running an “AI OS” stops sounding crazy once you see it working. I like Martin Mueller's isolation model here. The Agent Hub project I've been working on takes a similar approach, running agents in Docker sandboxes (lightweight VMs built for exactly this use case).

Scaling MCP adoption: Our reference architecture for simpler, safer and cheaper enterprise deployments of MCP
Cloudflare lays out a very real architecture for scaling MCP across an enterprise. This is exactly the kind of thing you have to think about. Context, auth, routing, and security all become first-class concerns once your agents start doing real work.

Multi-agent coordination patterns: Five approaches and when to use them
I’m a huge fan of patterns, and seeing these emerge in the agent ecosystem is exciting. We’re finally getting reusable, battle-tested approaches that unlock real capabilities instead of everyone rebuilding one-off hacks.

I Tested Three Spec-Driven AI Tools. Here’s My Honest Take.
I’m still not sold on spec-driven being the end state. It feels a bit too waterfall for how fast things move. That said, it’s a solid pattern for documentation and grounding the model upfront.

S3 Is Not a Filesystem (But Now There's One In Front of It)
Corey Quinn still says S3 isn't a filesystem, but he believes layering a real one in front of it actually makes sense. He also says the pricing is pretty reasonable for what you get, which means a lot coming from him.

The advisor strategy: Give Sonnet an intelligence boost with Opus
This is a really smart cost pattern. Let the expensive model do the hard thinking once, then have the cheaper models following instructions, which they're surprisingly good at.

Podcasts, Videos, and more

AWS Distinguished Eng: Learning From 3000 Incidents And How Engineering Is Changing | Marc Brooker
Marc Brooker shares some really interesting insights here on finding impactful engineering problems and system design patterns that actually hold up under real-world pressure. His advice for both junior and senior engineers navigating the AI shift is worth your time for a listen.

New from AWS

• Amazon CloudWatch Logs Insights now supports saved queries with parameters
• Spring AI SDK for Amazon Bedrock AgentCore is now Generally Available
• Amazon OpenSearch Serverless now supports Derived Source for storage optimization
• Amazon Bedrock Data Automation now supports custom vocabulary
• Amazon Redshift introduces key performance optimization for Top-K queries
• Amazon OpenSearch Service supports Managed Prometheus and agent tracing
• Amazon S3 Lifecycle pauses actions on objects that are unable to replicate
• Amazon CloudWatch pipelines now supports drop and conditional processing
• Amazon CloudWatch pipelines introduces new compliance and governance capabilities
• AWS Private CA now supports customer managed permissions for cross-account sharing
• AWS Billing and Cost Management Dashboards Now Supports Scheduled Email Delivery
• Amazon OpenSearch Serverless now supports Zstandard (zstd) codec for index compression
• Amazon RDS Blue/Green Deployments now supports Amazon RDS Proxy
• Oracle Database@AWS is now available in twelve AWS Regions

Final Thoughts 🤔

It’s getting a lot easier to break things.

Between models that can uncover vulnerabilities faster than ever, agents operating with increasing autonomy, and systems that were never designed for this kind of pressure, the attack surface is expanding in real time. The tools are getting smarter, but so are the ways they can be used against us.

And we’re still figuring out how to keep up with it.

But there are signs the industry is adjusting.

We’re seeing better visibility into AI spend, more structured ways to manage and govern agents, and even isolated environments becoming the default for running them. There’s a growing recognition that if agents are going to act on our behalf, we need systems that can track them, constrain them, and clean up after them when things go sideways.

Even S3 Files is part of that story. Not because it changes what S3 is, but because it makes powerful primitives easier to use, observe, and secure correctly.

It’s not clean yet, and it’s definitely not solved.

But the direction is clear. Smarter systems, stronger boundaries, and a shift toward building with the assumption that things will break.

Because they most definitely will.

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.