Weekly Review — Sep 7–13, 2026
written by Stefan Christoph
TL;DR: Five posts this week, split cleanly in two. Three build out the agentic web: why the content-API-versus-HTTP debate is a category error, and a tiny publisher whose HTTP 402 carries both a price and its license terms. Two turn to human judgment in the AI era: the full prompt-to-context-to-memory arc built on AWS, and a sharper career question, which of your work is actually verifiable. And one steps away from tech entirely, into what people in IT and media can do after an extremist party won 43.8 percent in a German state election.
Two distinct threads this week, and I would rather name them honestly than force one label over both. The agentic-web series kept going, following a single article through the layers that decide how an agent reaches it and what it is allowed to do afterward. Alongside it, two posts stepped back to ask what humans still own once the machine gets good: which part of the context actually matters, and which part of the work is truly ours to check. And one post sits deliberately apart from all of that.
This Week on the Blog
Why Doesn’t the Web Just Use HTTP? It Does, and It Doesn’t
The debate over dedicated content-ingest APIs versus plain HTTP is a category error. HTTP plus x402 already handle delivery and payment, while the IAB’s CoMP spec is a terms-and-marketplace layer that puts payment, licensing, and blocking explicitly out of scope and lists plain HTML as a valid delivery format. The real reason publishers reach for marketplaces is not a hole in the protocol, it is market power: collective bargaining, predictable revenue, and a legal paper trail.
Worried About Germany? You Should Be. What Can We Actually Do?
An unusual, personal post, written after the far-right AfD won 43.8 percent in the Saxony-Anhalt state election. It does not end with an answer, it ends with a question for people in IT and media, since the machinery that rewards outrage and simple stories is one our industry built. It lays out five research-backed starting points, from media literacy to prebunking to ranking audits, and asks for objections rather than applause.
Terms in the 402: The Payment Challenge That Also Tells You What You’re Licensed To Do
A tiny, public, dependency-free publisher whose HTTP 402 carries both a price and its license terms: an RSL license link in the response header plus a CoMP-style usage declaration at a well-known URI. An agent declares its intended use, reads the terms, refuses the uses the license prohibits, pays for the ones it permits, and gets the content. The whole thing runs locally on the Python standard library, and the code is on GitHub.
From Prompt to Context to Memory: The Arc, Grounded on AWS
Prompt engineering grew into context engineering, and memory is the rung after that. This walks the whole arc as a build on AWS: Bedrock prompt caching to reuse a stable context cheaply, Knowledge Bases with Titan embeddings for retrieval, AgentCore Memory for short and long term recall, and Guardrails at the boundary, with runnable code for the load-bearing steps. The recurring lesson is that better context is precision, not volume, and none of these services ships turned on.
Which of Your Work Is Actually Verifiable?
AI absorbs the work that is cheap to verify, which is the shared point of Jason Wei’s verifier’s rule and Christian Catalini’s Measurability Gap. Neither model accounts for the step before that, where you decide how verifiable a task even is, and a Zeynep Tufekci lecture argues we are consistently bad at it. The real career risk for a Solutions Architect is not that verifiable work gets automated, it is that we keep misjudging which of our work is verifiable, and wave the checking toward the machine only to find the checking was the job.
The Thread This Week
Four of the five circle one question from two directions: what decides what is allowed, and who is doing the deciding. On the agentic web it is literal, a 402 response and a license file settling what an agent may pay for and then do with the bytes. In the two AI-judgment posts it turns back on us, deciding what context a model should see and which of our work is actually ours to check. The Germany post sits deliberately outside that frame, and I left it there on purpose, because some weeks do not reduce to one neat theme and pretending otherwise would be the wrong move.
Further Reading
Things I read this week that didn’t get their own post, all public:
- Databricks: how we eliminated $1M/year of wasted AI agent spend Silent MCP tool failures made agents retry endlessly; tracing the tool calls surfaced seven small bugs worth roughly a million dollars a year in wasted tokens and lost productivity, and the fix took about an hour. A concrete argument for instrumenting agent tool calls the way you would any other hot path.
- Spotify’s Portal cut Claude Code token usage by 90% Route the I/O-heavy agent work, reading files and scanning code, to a cheap declarative model and reserve the frontier model for actual reasoning. A clean two-tier pattern for the teams already spending hundreds per developer per month on tokens.
- Kestra 2.0 exposes flows as MCP tools The orchestration platform now separates control and data planes and can expose its workflows directly as MCP tools for agents. A glimpse of where existing infrastructure meets the agentic layer without a rewrite.
- Pretraining progress is mostly data A controlled comparison attributes roughly a 12x compute-efficiency gain to data improvements against 3.7x from model-recipe changes. The uncomfortable takeaway for anyone chasing architecture: dataset curation is the bigger lever.
- Vortex: streaming columnar data from S3 to the GPU A zero-copy columnar format aimed at getting training data off object storage and onto the GPU without the usual conversion tax. Relevant if your ML training is quietly bottlenecked on data loading rather than compute.
- PostgreSQL 19’s WAIT FOR LSN: read-your-writes off replicas A primitive that turns replica consistency into a timeout budget, so you can read from a replica and still see your own recent write. A pragmatic middle ground between always hitting the primary and hoping the lag is small.
- Jitter is the cheapest reliability fix you are not using Adding randomness to retry and refresh intervals stops synchronized clients from stampeding a service in lockstep. One of those changes that costs a few lines and prevents a whole class of self-inflicted outages.
- Meta’s ZGateway: a proxy in front of ZippyDB Putting a gateway proxy ahead of the datastore cut connection fan-in by roughly 97 to 98 percent. A tidy case study in how a proxy layer tames the connection explosion that arrives with scale.
Until Next Sunday
Two of this week’s posts are really the same question pointed at different targets: what decides what is allowed, and who is doing the deciding. So here is mine for you: in your own work, which task have you quietly filed under “the machine can check that,” and are you sure the checking is not the part that was actually yours? Tell me where you drew that line.
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)