Off-by-none: Issue #371

July 7, 2026

Could your SaaS be replaced by a Markdown File? 📝

In our previous issue, Anthropic launched Claude Sonnet 5, CloudFormation got much faster, and OpenAI started making Jalapeños. In this issue, Claude Cowork breaks free of your laptop, MiniMax lands on Bedrock, and we weigh what taste and judgment are still worth. Plus, we have lots of awesome content from the cloud, serverless, and AI communities.

News & Announcements

Most of the interesting news from this week is around agent plumbing, and AWS still seems to be doing a lot of that work. The most interesting is structured memory filtering with metadata in AgentCore Memory, which layers attribute-based filtering on top of namespace isolation. Memory scope and attributes are two different things, and it's nice to see that AWS is getting this one right. AgentCore also bumped its default runtime quotas, now up to 200 agent interactions and 25 new sessions per second, with US East and West carrying 5,000 concurrent sessions. More headroom is good, just be sure to use it wisely.

On the model side, MiniMax models are now on Amazon Bedrock across 14 regions. I like the MiniMax models, and with tool-calling, implicit prompt caching, and $0.30 price tag per million input tokens, it could be a nice drop-in for your agentic workflows. AWS also shipped an open source MCP server for the Registry of Open Data, giving AI assistants access to more than 1,100 public datasets for natural-language discovery. Useful for anyone who wants to ground their research against satellite imagery, climate, genomics data, etc. without knowing the catalog cold.

For the folks who actually keep things running in the cloud, ECS picked up real-time deployment observability in the console with a live timeline, circuit breaker monitoring, and failed-task diagnostics at no extra charge, which is exactly the kind of thing you appreciate at 2 a.m. while watching a rollout stall. And Cognito moved API rate limits to self-service, so you can raise them from the console up to your account maximum instead of opening a Service Quotas ticket and waiting.

Outside of AWS, Claude Cowork landed on web and mobile with background execution and scheduled runs, so a task can prep your morning briefing at 6am and ping your phone when it needs a decision before continuing. That was a big theme at the AI Engineer World's Fair last week, workflows moving off your laptop so you're no longer limited by compute (just money). Together AI raised $800M at an $8.3B valuation to grow its AI-optimized public cloud. They claim their ATLAS technology speeds up some inference workloads by as much as 400%.

Vercel will now run any Dockerfile on its Fluid compute platform, so Go, Rails, or Spring Boot deploy with the same preview-and-scale workflow as everything else. And EventCatalog v4 reframes itself from event docs to a broader architecture catalog, adding a Systems resource type and an agent that keeps your docs synced to the codebase.

Tutorials

• Teaching models to forget: Selective unlearning with Amazon Nova by Qian Hu
• Build a serverless image editing agent with Amazon Bedrock AgentCore harness by Salman Ahmed
• Automatically redact PII in images with Amazon Nova by Caroline Des Rochers
• Data masking in Amazon RDS for Oracle by Jobin Joseph
• Choosing a Claude model and effort level in Claude Code by Lydia Hallie
• I Built PR Preview Environments With AWS Lambda MicroVMs and Cut Staging Costs by 78% by Jatin Mehrotra

Reads

Ten years of micro-frontend decisions, condensed into a skill by Luca Mezzalira
Luca took a decade of hard-won micro-frontend lessons and packed them into a skill that keeps AI agents from wrecking your boundaries. If you've ever watched a cross-boundary import or a shared global sneak into a codebase, this is the guardrail you wish you'd had.

What America has meant to me by Shawn "swyx" Wang
Swyx walks through his three "runs" in America across almost twenty years, from finance to engineering to founding Latent Space and the AI Engineer movement.

Claude Sonnet 5 Launch Analysis: What Changed, What Matters, and What to Validate by Guille Ojeda
Guille breaks down what actually changed in Sonnet 5, including the adaptive thinking defaults, effort controls, and a tokenizer shift that bumps your token counts by about 30%. The real takeaway is to validate against your own workloads instead of trusting the benchmarks.

In defense of AI mandates by Charity Majors
Charity argues that top-down AI mandates give managers the cover they need to actually spend time and budget on adoption. Her point is that companies have to decide whether AI is existential or optional, then fund that decision like they mean it.

A Field Guide to Claude Fable: Finding Your Unknowns by Thariq Shihipar
Thariq lays out a framework for categorizing what you don't know and applying the right pattern at each stage of an AI coding workflow. The blind spot passes and pre-implementation planning are the parts you should steal first.

I replaced my GitHub runners with Lambda MicroVMs, and maybe you should too by Luc van Donkersgoed
Luc swaps GitHub-hosted runners for Lambda MicroVMs and is honest about the tradeoffs, which land at minimal cost savings but a few minutes faster per run. The real cost is the operational overhead of owning your runner infrastructure, and he doesn't pretend otherwise.

What if building an AI chatbot was as easy as snapping LEGO bricks together? - AWS Blocks by Luis Fernando de León Ramírez
Luis shows off AWS Blocks, the new AWS open-source framework that turns TypeScript into AWS serverless services with a snap-together feel. The walkthrough covers local dev, sandbox testing, IAM, and shipping a Bedrock Nova Lite chatbot with real code.

Lambda MicroVMs vs AgentCore Runtime: When to Use Each for Production Agents by Gerardo Arroyo
Gerardo lays out when to reach for Lambda MicroVMs versus AgentCore Runtime for production agents, and where the two actually complement each other. The framing around coding agents that need secure execution sandboxes is the useful bit.

15 things I learned at AI Engineer World’s Fair 2026 by Dave Thackeray
Dave's fifteen takeaways cover the mismatch between probabilistic models and deterministic infra, context cost, and patterns like semantic routing and post-generation veto systems. Since I was there too, I can tell you his list holds up.

The AI Chatbot Era Is Ending. Teams Are Optimizing the Wrong Layer. by Tyler Folkman
Tyler Folkman argues that teams should stop optimizing prompts and start designing delegation frameworks for AI agents. He presents a five-layer Delegation Stack based on hundreds of agent sessions, backed by recent data from OpenAI and Anthropic showing enterprise shifts toward agent-based workflows.

Podcasts, Videos, and more

Sonnet 5 review: I ran 64 generations to find out if it's worth it
Claire ran 64 generations pitting Sonnet 5 against Sonnet 4.6, Opus 4.8, GPT-5.5, and Gemini 3 Pro with her own eval framework. She scored PRD quality, prototype generation, agentic completion, and agent personality, and the results surprised her.

AI Engineer - YouTube
The AI Engineer channel is where the World's Fair sessions land once they're posted, so subscribe if you couldn't make it in person. Right now, all the livestreams of the keynotes and main stage track are posted, so check those out when you get some time.

New from AWS

• AWS Security Hub adds impact analysis for exposure findings
• AWS Secrets Manager adds managed external secrets support for Paddle and GitLab
• Amazon CloudWatch supports creating alarms from log queries
• CloudWatch Application Signals now supports Dynamic Instrumentation
• AWS Config now supports 8 new resource types
• Amazon ECS Express Mode now supports custom task definitions
• Amazon OpenSearch Service optimized for log analytics
• Amazon ECS now supports configurable deployment circuit breaker settings
• Amazon EMR Serverless now supports larger worker sizes to run more compute and memory intensive workloads
• Amazon EKS Auto Mode reduces GPU management fees by up to 60%
• Amazon ECS Managed Instances reduces GPU management fees by up to 60%

Thoughts from Social

“Half of the companies here at @aiDotEngineer could be a markdown file!” ~ @theo 🌶️🌶️🌶️ pic.twitter.com/kxaTCZUd2G

— Jeremy Daly (@jeremy_daly) July 2, 2026

Developer Tools

Simplify model selection in Amazon Bedrock with the open source Model Profiler by Maria Oliva Calero
Maria walks through the Bedrock Model Profiler, an open-source tool that pulls 120+ foundation models into one searchable view. The serverless pipeline stitches together five AWS APIs and two public sources so you can filter on pricing, region, quotas, and lifecycle without tab-hopping.

Final Thoughts 🤔

Theo's comment that half the companies at the AI Engineer World's Fair could be a markdown file landed hard for some. Because he's not entirely wrong. Point a capable model at a well-structured markdown file, wrap it in a skill and hand it to Claude, and a lot of what those startups demoed just falls out the other end. Instruction-following has gotten good enough that the markdown file basically becomes the product.

The catch is that a markdown file captures what you want done, not the judgment behind how it's done. Romain Huet from OpenAI made this point in his keynote: engineering has always been about solving problems by combining the latest science "with design, with taste, with judgment, and most of all, imagination" to make something people can actually use. That judgment is what a lot of SaaS companies are actually selling. They've spent years sequencing business processes and getting them battle-tested across thousands of customers, and the good startups are staffed by domain experts who already worked out the right way to handle a problem so you don't have to. Replace that with a markdown file and you're trading years of accumulated judgment for something you wrote in an afternoon.

Then there's the bill. Running a skill once is cheap and kind of magical. Running it thousands of times a day against a frontier model is a line item nobody budgeted for. The workflows that actually scale are mostly deterministic, with the model reserved for the handful of points where a real decision has to be made: read this data, decide whether to proceed, escalate to a human, or route somewhere else. That's where an LLM earns its cost. Wrapping the whole process in one because you can is how you light money on fire.

And almost all of it shows up in the last 10%. AI might get you 90% of the way to a working product in a weekend, and that first 90% is impressive. But the last 10% is the edge cases, the industry quirks, and the hard-won defaults that come from domain experts and a team of PMs and engineers talking to real customers. You'll build the features you use every day and feel great about it, right up until you hit the edge cases and nuances that really matter. That gap is brutal, and no markdown file is going to close it for you.

So could your SaaS be replaced by a markdown file? For a thin wrapper, sure, and probably very soon. For anything that encodes real judgment about a hard problem, the markdown file gets you a convincing demo and a bill for the 10% you didn't build. Know which one you're building.

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.