Weekly Review — June 15-21, 2026
written by Stefan Christoph
This is the Weekly Review, a Sunday digest of everything that went up on the blog this week, plus a short list of things I read but didn’t write about. If you only have ten minutes on a Sunday, this is the one to read.
This Week on the Blog
MCP Strategies on AWS, Part 1: Reading the Whole Guide
AWS published a Prescriptive Guidance paper that organizes Model Context Protocol into three pillars — tool design, hosting, and governance — and this post is the map of the whole thing. It pulls out the parts that are actually actionable (keep tools to eight parameters or fewer, bundle any three-plus-call workflow into one tool, split a server past fifty tools, and remember each tool definition costs 250-500 tokens on every model call) and lines the guidance up against the security concerns the wider MCP community has been raising. It’s a guide-tier document that tells you how to think, not which buttons to click — so this kicks off a series that turns each pillar into runnable code.
AI Content Pipeline Deep Dive (5/5): Publishing
The final part of the pipeline series tackles the last mile: the nine mechanical steps between “done writing” and “live on the internet” that quietly kill most people’s publishing cadence. It shows the actual constraint files that drive image generation from performance data, LinkedIn teasers written as hooks rather than summaries, and a one-command deploy that handles build, deploy, and verification — compressing 30-45 minutes of tedious work into about two minutes of human review.
From Columbo to Coworker
Running several interactive agents at once feels like Lieutenant Columbo lives in your terminal (every few minutes one turns around with “just one more thing”), and this post reframes that as a management problem, not a tooling one. Instead of being the scheduler who gets interrupted all day, you run agents like a team with four ceremonies: a morning standup, grooming and delegation, office hours for blockers, and an evening collection. It’s an honest mid-journey report twelve days in: the headless half works better than expected, the interactive half still interrupts more than a good coworker would.
MCP Strategies on AWS, Part 2: Tool Design in Code
Part two takes the paper’s most checkable claims and writes code to test them. A token-tax counter on a realistic 20-tool GitHub server shows minimal definitions cost ~92 tokens each, but the paper’s recommended enriched form (output schemas, concrete examples, prompt-style descriptions) lands at 346 tokens each — squarely inside the paper’s 250-500 band, which turns out to be the cost of good tool definitions you pay on every call. A second demo shows the granularity trade-off: three granular tools force three model round-trips where one coarse workflow tool does the same backend work in one.
Your Voice Clone Is Only as Good as the Reference Clip
A voice clone inherits the cadence, pitch, and timbre of its reference clip, so the single biggest lever on quality is the clip you record, not the model — and almost nobody measures it. This is a build log on Voice Sample Studio, a local web app that records takes, scores each on acoustic quality and delivery, gives a keep/review/reject verdict with a star rating, transcribes it, and exports a clone-ready 24 kHz mono clip plus its reference transcript. The scoring runs entirely on CPU; two optional features tap the cloud for richer advice and a voice preview and degrade gracefully without it.
MCP Strategies on AWS, Part 3: Hosting, from Local to AWS
Part three walks the hosting ladder with code: a local stdio server with no auth, then a real remote deploy, then the managed-gateway option. The remote rung is the concrete one — a Lambda Function URL with AWS_IAM authorization, which bounds the server to a single AWS account because every request must be SigV4-signed by an in-account principal and an unsigned call gets a 403. It was deployed to a real account, the tools listed and invoked, the 403 confirmed, then torn down, with the actual output and a trap-based teardown that runs even if the process is killed.
The Thread This Week
Three of these posts are one series, and the series is the week’s whole argument in miniature: a strategy paper is only as useful as the code you can run against it, so each pillar gets turned into something measurable — token counts you can reproduce, a deploy you can sign and tear down. The other three rhyme with it. Publishing automates the craft around the writing; the Columbo post structures the human attention around the agents; the voice studio measures the input before the model. None of them are about a bigger model. All of them are about the discipline that decides whether the model’s output is any good.
Further Reading
Things I read this week that didn’t get their own post. All public:
- Who Operates Is the Moat — Analysis of four agent-orchestration startups converging on the same governance spine (governed work-item, approval gate, immutable audit). The real differentiator isn’t agent cleverness; it’s who holds the pager. A sharp companion to this week’s MCP governance pillar.
- Matt Wood — What The Garden Is For — A gardening metaphor for long-running agents: as they work autonomously for days, the human role concentrates toward judgment and direction rather than shrinking. Reads as the thoughtful counterpart to the Columbo post.
- Spec-Driven Development, One Year On (Boeckeler & Tacho) — Birgitta Boeckeler and Laura Tacho on where spec-driven development actually stands in 2026: one workflow for all sizes, reviewing markdown vs code, and whether spec-as-source is repeating old mistakes.
- AI-Generated Architecture Diagrams (Charly Wargnier) — Generating architecture diagrams from text instead of manually dragging boxes. A small, practical workflow win for anyone who still hand-builds diagrams.
- Ponytail — Multi-IDE Agent Plugin Framework — An open framework for running the same agent plugin across multiple IDEs. Worth a look if you’re tired of re-implementing agent tooling per editor.
- Team OKRs in Action (Paulo Caroli, martinfowler.com) — The argument that top-down OKR cascades kill commitment, and that high-performing teams define their own OKRs in conversation with leadership. A useful non-AI palate cleanser on how autonomy actually gets built.
- AWS Summit Hamburg Takeaways (Markus Großmann) — An external field read on the Summit: agentic AI dominates and the tooling has matured, but trust — not technology — is still the unresolved blocker for full autonomy.
Until Next Sunday
That’s the week. The through-line (value comes from the craft and measurement around the model, not from reaching for a bigger one) is the same one I keep landing on, and the MCP series is the clearest proof of it yet: a strategy paper only earns its keep once you can run code against it.
Which of these would you have led with? And what did you read this week that I should have?
About the Author
Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.
This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.
❤️ Created with the support of AI (Kiro)