Weekly Review — Sep 14–20, 2026
written by Stefan Christoph
TL;DR: Four posts this week, and one quiet habit running under all of them: taking something that looks new, mysterious, or personal and renaming it as something familiar and structural. Agent authorization turns out to be OAuth you already know. A suction cup is held on by the weight of the air. A recruiter’s flat manner is saturation, not a verdict. And the agentic web is not one business model winning but two, splitting by content type.
A good week is one where the answer to “what is really going on here?” keeps landing on something you already understood, just wearing new clothes. That was this week. Two posts pulled apart the plumbing of the agentic web, two looked at the physics and psychology of everyday life, and all four did the same move: strip off the scary label and point at the ordinary mechanism underneath.
This Week on the Blog
Foundations of Agent Authorization
Agent identity feels like a new frontier, but almost all of it is OAuth 2.0 and OpenID Connect, the same flows that already sign you into web apps. The post follows a single token through three shapes (a user delegating access, a service acting as itself, and a token exchanged to act on someone’s behalf) and shows that MCP and A2A reuse standard web auth rather than inventing their own. It closes by mapping every piece onto Amazon Cognito and Amazon Bedrock AgentCore Identity, so the whole chain has a concrete home.
Why Suction Cups Stick (It’s Not Suction)
The first Lunch Break Physics: nothing pulls a suction cup onto the wall. You live at the bottom of an ocean of air that pushes on every surface at about 10 newtons per square centimetre. Pressing the cup drops the pressure underneath, the atmosphere keeps pushing at full strength, and that imbalance pins it in place, with the holding force just the pressure difference times the sealed area. The same trick shows up in straws, breathing, and an octopus arm stuck to a dinner plate.
Asymmetry of Stakes: Staying Humane at Volume
To the recruiter it is interview number forty this month; to you it is the next five years of your life, and both readings are correct. The post walks through the research showing that compassion fades as numbers rise, that some of the numbing is a deliberate self-defense, and that the fix is not “feel more” but “feel differently.” It ends with a concrete playbook for both sides of the table, from the person processing the queue to the person who is case forty.
Who Pays for the Agentic Web, the Agent or the Advertiser?
The capstone that ties two series together. When the visitor is an AI agent, two opposite business models form: the agent pays for access through an x402 micropayment, or the advertiser pays to influence the answer the agent gives. The likely end state is not one winning but segmentation, tolls on must-have and timely content, ads on evergreen and authority content, the same split broadcast television settled on decades ago with carriage fees and ad sales.
The Thread This Week
Put the four side by side and the pattern is hard to miss. Each one takes a thing that looks novel or intimidating and hands you the familiar machine inside it. Agent auth is web auth. Suction is atmospheric pressure. A cold-seeming professional is a saturated one. A brand-new web economy is an old dual-rail media economy. Naming the mechanism correctly is most of the work, because once you can point at it, you know which lever to pull.
Further Reading
Things I read this week that did not get their own post, all public:
- Implementing defense-in-depth authorization for MCP tools on Amazon Quick: The natural sequel to the agent-authorization post: a valid token proves who you are, not what you may do. This walks a four-gate pattern (MFA, country geo-fence, group RBAC, and tool-level permission) enforced through an AgentCore Gateway interceptor.
- A shared agentic platform for Wood Mackenzie on Amazon Bedrock AgentCore: Wood Mackenzie reports that 88 percent of their AI proofs-of-concept never reached deployment, and blames architecture rather than the model. Their answer, APEX, is one shared platform handling runtime, identity, gateway, memory, and observability so teams ship business logic instead of rebuilding plumbing.
- How MRH Trowe enabled secure self-service AI agents in financial services: A German insurance broker gave roughly 400 employees governed AI agents in the first production month, combining Strands Agents, AgentCore, and LibreChat at about $14 per seat, with everything kept in the Frankfurt Region. A concrete regulated-industry reference for the EMEA sovereignty conversation.
- Run open weight models on Amazon Bedrock in AWS European Sovereign Cloud: Gemma 4 (Apache 2.0) is generally available on the Bedrock next-generation inference engine in the AWS European Sovereign Cloud, the first open-weight family in the sovereign region, with data staying in eusc-de-east-1. Useful if a customer needs Bedrock but could not use non-EU inference.
- Beyond the model: engineering AI infra with scientific judgement (Airbnb): Airbnb’s Insight Miner encodes the scientific method itself as infrastructure around the model, so an agent’s analysis of unstructured text is reproducible, auditable, and challengeable rather than just polished. The point that stuck with me: the method is as much the product as the answer.
- How Claude’s text watermark works (Anthropic): A clear explainer of the SynthID-based watermark now added for EU AI Act compliance. It changes only the source of randomness in low-stakes word choices, so there is no quality or cost impact, it survives light editing but not a full rewrite, and generated files also carry C2PA provenance metadata.
Until Next Sunday
The habit that connected this week was renaming: every post replaced a mysterious label with a plain mechanism. So here is the question I would love an answer to. What is one thing in your own work that everyone treats as new and scary, but is really an old mechanism in a new costume? Reply and tell me which one you keep having to un-mystify for people.
This is the Weekly Review, and it also goes out as Sunday’s newsletter.
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)