Weekly Review — July 27-August 2, 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.
Prefer to watch? This narrated walkthrough covers the whole digest in a couple of minutes.
This Week on the Blog
Tesler’s Law: You Can’t Delete Complexity, Only Choose Who Owns It
Tesler’s Law, the conservation of complexity, says a system has an inherent amount of complexity that can be reallocated between owners but never deleted. The useful move is to split the number most people conflate: the total complexity of the whole system, and the complexity you personally have to manage and operate. A well-run managed service shrinks what you manage, which is a genuine win; AI now makes it cheap to build a focused component with exactly the features you need and less inherent complexity, which can lower the total, but you own and operate it, so what you manage goes up. Buy less to manage, build with AI for less total but you hold the pager. Third and final in “Laws & Disorder.”
Generate-and-Verify: What P vs NP Teaches Us About AI Reasoning
The defining feature of the complexity class NP is that a candidate answer is cheap to check even when finding one is brutally hard, and that same asymmetry explains a lot of how modern AI reasoning works. Test-time compute spends more effort on the search, a verifier or reward model or unit test plays the cheap checker, and an agent loop is just propose-then-verify repeated. In compute-matched tests, a small reliable verifier gating a strong generator can beat scaling the generator alone — but the trick only pays when a cheap, trustworthy checker exists, a wrong verifier is worse than none, and none of this means a model “solves” NP-hard problems. It searches, and lets a cheap check do the confirming.
Understanding Is the New Bottleneck
When agents write the code, the hard part is no longer writing it — it is understanding what the agent wrote well enough to keep steering. Geoffrey Litt’s framing is that the human job shifts from verifying correctness, which agents increasingly do themselves, to understanding so you can participate and have the next idea; skip it long enough and you accrue “cognitive debt.” His fix borrows from teaching: have the agent write an explainer for its own change and quiz you on it before you ship. I ported his ExplainDiff idea to a Kiro CLI skill and ran it on a real diff — it produced a self-contained explainer with a five-question quiz, and the rule that comes with it is a good one: don’t send a PR you can’t pass the quiz on. The code is on GitHub.
The Thread This Week
Read together, the week is about where the human stays in charge as agents take on more. Tesler’s Law says you can’t delete complexity, only decide who owns it — a judgment call, not a technical one. Generate-and-verify says the durable human asset is the checker: own a cheap, trustworthy verifier and let the model search. And “understanding is the new bottleneck” says the role worth protecting is comprehension, because that’s what lets you have the next idea. Judgment, ownership, understanding — the parts that concentrate rather than shrink.
Further Reading
Things I read recently that didn’t get their own post, all public:
- Steering: mid-turn course-correction in Kiro CLI (Doug Clauson) — When an agent heads the wrong way, you can now steer it at the next tool boundary without stopping the run. A concrete companion to the generate-and-verify loop: cheap correction beats an expensive restart.
- A good paper on self-improving agents (Elvis Saravia) — A pointer to research on agents that get better over their own runs. Relevant to the propose-then-verify framing when the verifier feeds back into the next attempt.
- What the garden is for (Matt Wood) — A gardening metaphor for long-running agents: as they work autonomously for days, the human role concentrates toward judgment and direction rather than shrinking. The clearest companion to “understanding is the new bottleneck.”
- Books at the AI intersection (Constantin Gonzalez) — A curated set of unexpected book picks that broaden how you think about AI beyond ML textbooks. Good fuel for the “have the next idea” muscle.
- Loom for AWS — a secure agent platform (AWS Open Source Blog) — An enterprise agent platform on Strands and Bedrock AgentCore Runtime with RBAC/ABAC, tag-based governance, and human-in-the-loop patterns. The operational side of keeping humans in the loop as agents scale.
- Agentic AI with Amazon Bedrock (Tracy Geist) — Enterprise adoption patterns for agents built on Bedrock’s managed agentic infrastructure. Useful context for where the build-vs-buy line from the Tesler’s Law post is actually being drawn.
Until Next Sunday
That’s the week, and the close of “Laws & Disorder”: a law that settles build-vs-buy by asking who owns the complexity, an NP asymmetry that tells you to spend on the search and cheap-check the result, and a reminder that understanding, not writing, is now the thing to protect. The thread is the human role concentrating toward judgment, ownership, and comprehension. Which of those three are you actively defending in your own work?
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)