Off-by-none: Issue #366

May 19, 2026

The Flat-Rate Honeymoon is Over 📈

In our previous issue, Claude Platform set up shop on AWS, ElastiCache learned to do full-text and hybrid search, and Ampt rolled out Node.js 24 as the default runtime. This week, Anthropic brings subscription clarity to Claude, Codex goes mobile, and Amazon DSQL gets CDC on DPUs. Plus, we've got plenty of awesome content from cloud, serverless, and AI communities.

News & Announcements

Anthropic announced this past week that paid Claude plans will get a dedicated monthly credit for programmatic usage starting June 15. Pro gets $20 in monthly credits, Max 20x gets $200, and anything past the allocation rolls onto API rates. If you've been running SDK loops, claude -p jobs, or GitHub Actions agents on your subscription, the honeymoon is over. Theo Browne took it about as well as you'd expect. His reaction video is appropriately titled "I'm done," and his X post promised to donate $10 to open source for every screenshot of a cancelled Claude Code plan that was shared. He's not wrong to be frustrated. The rules have been ambiguous for quite some time, and this announcement does provide some clarity, just not what most of us were hoping for. The bigger lesson here is about AI platform risk. Cancelling Claude Code doesn't fix the real problem, and if your business depends on one vendor's pricing staying frozen forever, your subscription isn't the thing that needs changing.

Even with all the developer backlash, Anthropic seems undeterred. They continue to ship more and more enterprise plumbing. Claude Managed Agents now support self-hosted sandboxes and MCP tunnels, so your agents' tools can run inside your own infrastructure while orchestration stays on Anthropic's platform. Cloudflare, Daytona, Modal, and Vercel are all in the launch lineup, with Cloudflare getting its own first-class slot including Browser Run for automation and quick-start templates. Anthropic also pushed two vertical packages: Claude for the legal industry and Claude for Small Business. Legal feels like the obvious play. High-value industries with messy document workflows are where managed agents make a lot of sense, so long as it doesn't keep hallucinating case law.

Elsewhere in the agent space, OpenAI brought Codex to the ChatGPT mobile app with Remote SSH now GA, programmatic access tokens, and HIPAA compliance for healthcare. Coding from your phone still sounds like a stretch, but the strategy is right: meet developers wherever they happen to be. Also, Temporal added Workflow Streams and Standalone Activities to its durable execution platform, both aimed squarely at the AI-in-production folks. Durability and debuggability are the two things most agentic systems are starving for, so the direction makes sense.

AWS had a busy week too. Amazon Aurora DSQL now supports change data capture in preview, streaming database changes to Kinesis Data Streams for event-driven apps and real-time analytics. I think the Distributed Processing Units plus Kinesis pricing is going to trip a few people up, but I've been wrong before. Amazon Bedrock launched Advanced Prompt Optimization (with a deeper writeup on the AWS blog), automating prompt comparison across up to 5 models with custom evaluation metrics, Lambda-based scoring, LLM-as-a-Judge rubrics, and multimodal inputs including images and PDFs.

On the Lambda side, AWS added scheduled scaling for functions on Lambda Managed Instances via EventBridge Scheduler, useful for adjusting capacity ahead of expected traffic, and ARC Region switch now automates Lambda event source mapping execution during failovers across Kinesis, DynamoDB Streams, MSK, and SQS, with cross-account support. That last one is very cool. CloudFront got two updates: Passthrough Mode for mTLS that forwards certificates to origins without edge validation, and configurable usage allowances on the Premium flat-rate plan from 500 million to 6 billion requests and 50 TB to 600 TB per month. And EventBridge Scheduler added 619 new SDK API actions across 13 services, bringing the total coverage to over 270 AWS services.

Finally, on the security side, Wiz's Runtime Sensor for Google Cloud Run is now GA, with 2000+ detection rules and AI-driven investigation through their Blue Agent. Serverless container monitoring built for serverless containers. Who would have thought?

Tutorials

• Zero-downtime DynamoDB construct migration: from Table to TableV2 with cdk orphan by Lee Hannigan
• Getting started with Change Data Capture in Amazon Aurora DSQL by Vijay Karumajji
• Dynamic Looping Comes to AWS SAM by Eric Johnson
• Best practices for computer and browser use with Claude by Lucas Gonzalez and Luca Weihs
• AtMostOncePerRetry vs AtLeastOncePerRetry Semantics in Lambda Durable Function Step by Rishi
• Build custom code-based evaluators in Amazon Bedrock AgentCore by Bharathi Srinivasan
• Layered Configuration in Claude Code by Michael Walmsley
• Per-tenant DynamoDB isolation with the Token Vending Machine pattern by Monica Colangelo
• Lambda Durable Functions, When You Don't Need Step Functions by Lewis Sawe
• Live Canary Deployments with AWS SAM, the New WebSocket API Resource, and Lambda Durable Functions by Darryl Ruggles

Reads

The founder's playbook: Building an AI-native startup
Anthropic walks through Idea, MVP, Launch, and Scale for AI-native startups, with real founder stories woven in throughout. Playbooks aren't the whole answer, but if you're staring at a blank canvas, this is a pretty good starting point.

How Claude Code works in large codebases: Best practices and where to start by
A complete walkthrough of Claude Code's extension points (CLAUDE.md files, LSP integrations, MCP servers, subagents) and how they shape behavior in enterprise codebases. If you're getting mediocre results from Claude Code once you push past your toy project, this will probably help you fill some gaps.

Project Glasswing: what Mythos showed us by Grant Bourzikas
Cloudflare tested Anthropic's Mythos Preview model on a number of their repos and got firsthand knowledge of why off-the-shelf coding agents fall short. The multi-stage architecture they built is a useful reference for anyone doing serious agentic work outside of the basic coding use case.

Opinion | The Generation That Grew Up With A.I. Hates It by Michelle Goldberg
Only 18% of Gen Z is hopeful about AI, and 47% of voters under 30 rate it as mostly bad. As someone with two daughters in that demographic, I can't say I'm surprised. They've watched the technology arrive with lots of promises but not a lot of upside for them.

Local agents scare me by Allen Helton
Allen walks through four attack vectors for local AI agents (shared userland, network adjacency, poisoned context, persistent state) and makes the case that traditional IAM controls don't fit. Definitely worth reading before you give any agent unrestricted shell access on your machine.

Is AWS Lambda Tenant Isolation Mode Enough for SaaS? by Ran Isenberg
Ran breaks down what Lambda's tenant isolation actually solves and what it doesn't. The compute side is handled, but data access control is still on you, which has always been the hard part of multi-tenancy.

10 Practical Serverless Architecture Lessons from AWS Summit London 2026 by Siddarth Patil
A grab-bag of serverless patterns from AWS Summit London: Lambda boundaries, async with EventBridge and SQS, cold starts, cost management, and applying the same patterns to GenAI workloads. Most of it is table stakes if you've been doing this a while, but the GenAI section is worth a skim.

Cross-Domain Governance by Aaron Sempf
Aaron explains how autonomous systems should behave when they cross organizational boundaries, proposing monotonic reduction: authority can only be restricted, never amplified, as you move outward. It's a tidy way to think about a problem most agent platforms haven't even acknowledged yet. I always feel smarter after reading his stuff.

Podcasts, Videos, and more

Building Apps with AI + MCP Servers | Serverless Office Hours
Brian Zambrano joins Darko Mesaroš to build a serverless application from prompts using Kiro and MCP servers. A solid walkthrough from natural language to deployed AWS infrastructure.

How I AI: HTML is the new Markdown: How Anthropic engineers are building with Claude Code
Claire Vo interviews Thariq Shihipar from Anthropic's Claude Code team on the shift from Markdown to HTML for AI output, plus patterns like living design systems and micro-apps. I'm all for the HTML move. Markdown files have gotten easier and easier to gloss over, and giving the model a proper display layer makes the output feel a lot less disposable.

Serverless CrAIc Ep84 AI-Generated Code Is a Liability: Technical Debt & Engineering Excellence
The Serverless Craic crew digs into the velocity versus debt tradeoff in AI-generated code, including the awkward truth that more tests and more code don’t always mean better quality. The discussion around engineering excellence as a counterweight to AI-driven throughput is where this one gets especially heady. They argue that production code still has to be maintained by humans eventually. Let's hope.

New from AWS

• Reference stack outputs across accounts and Regions with AWS CloudFormation and CDK
• AWS Transform agents now available in Kiro, Claude, Cursor, and Codex
• AWS Transform adds agentic AI assistant to the AWS Toolkit for Visual Studio
• Announcing general availability of Amazon EC2 M3 Ultra Mac instances
• Amazon Redshift adds ALTER TABLE for Iceberg tables and writes via the AWS Glue Data Catalog mount
• Amazon EMR Serverless is now available in additional AWS Regions
• AWS announces AWS Interconnect - multicloud connectivity with Oracle Cloud Infrastructure in preview
• Amazon CloudWatch Logs announces increased query result limits
• AWS Organizations now supports higher quotas for service control policies (SCPs)
• AWS Security Agent now supports full repository code reviews
• Amazon Route 53 Domains adds support for 34 new Top Level Domains including .app, .dev, and .health.
• AWS SAM CLI adds AWS CloudFormation Language Extensions support to accelerate local serverless development

Final Thoughts 🤔

There's a technical nuance to this Claude Code change that's worth pulling apart. Running claude -p isn't the same as hitting the API directly. Claude Code ships with caching, tool-use optimizations, context management, and prompt structuring that make the interactive product feel as useful as it does. When you wire claude -p into a script, you get those same optimizations applied to your automated workflows. Hit the raw API yourself and you're rebuilding all of that from scratch, usually badly, burning extra tokens on every loop, and probably wrecking the economics in the process.

That's why this change stings more than a normal pricing adjustment. I understand metering for large-scale programmatic use, but the bigger shift is that now the cheap, optimized path is effectively reserved for interactive use. If you want automation, you either pay API rates or stay inside one of Claude's tightly controlled (typically not great) interfaces.

That's the part that bothers me. You're paying for more than just access to a raw model. You're paying for the orchestration layer around it: the caching, context handling, tool execution, prompt shaping, and all the little optimizations that make Claude Code actually useful day to day. Whether those requests originate from a human typing into a terminal or a script running in the background seems mostly immaterial.

And that's the bigger question this raises for the industry. Are these systems ultimately meant to become programmable infrastructure, or are they meant to remain interactive products with a human sitting in front of them? Because the economics matter. Automation only works when the cost structure makes sense. If the optimized path is reserved for interactive use while automated use is pushed onto significantly more expensive APIs, then we're implicitly putting limits on how far these tools can evolve beyond "copilot" workflows.

That's worth thinking about before we build entire engineering organizations around them.

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.