Weekly Review — July 13-19, 2026

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

Conway’s Law: Your Org Chart Is Your Real Architecture

Conway’s Law says a system ends up mirroring the communication structure of the organization that built it — so on AWS, as a rule of thumb rather than an iron law, your service and account boundaries drift toward your team boundaries unless you actively counter it. The classic counter-move is the Inverse Conway Maneuver: shape the teams first to nudge the architecture toward what you want, and let the law work for you. The 2026 twist is the bet that AI coding assistants lower the cross-team coordination cost the law runs on, and teams are starting to slot fleets of agents into the org chart too — whether Conway’s mapping loosens, replicates faster, or even applies to agents at all is the open question worth sitting with. This kicks off “Laws & Disorder,” a short series on the engineering laws we half-quote.

Why AI Tokens Are So Expensive

An afternoon argument over café napkins in Austria turned into this: LLM generation has two phases with different bottlenecks — prefill processes the prompt in parallel and is usually compute-bound, while decode emits tokens one at a time and is usually limited by memory bandwidth. Long contexts grow the KV cache, which — depending on batch size, precision, and serving configuration — eats accelerator memory that could otherwise hold more concurrent requests, and resending full chat history makes the cumulative input tokens across a conversation grow quadratically with the turns. Amazon Bedrock prompt caching does not repeal that arithmetic; it reuses matching prefix state at a reduced read rate, which AWS reports as up to 90% lower input-token cost for supported workloads.

From Robot Soccer to the Agentic World Cup

Twenty years ago a year-long university project built three soccer-playing cube robots with one overhead camera and a shared world view; this year AWS runs a four-hour workshop where you build five autonomous Amazon Bedrock agents to play the same game. The tools changed completely, and the environments differ in every physical detail — but the coordination patterns did not: roles, shared state, and time-critical decisions between agents. My favorite lesson came from “Bob,” the third robot whose job was mostly to stay out of the way — an early lesson in coordination through restraint (assigned, not emergent: staying out of the way was Bob’s role), where an agent’s value is sometimes what it does not do.

The Thread This Week

Read together, the week is three views of one idea: the structure you don’t design still gets drawn. Conway’s Law draws it in your org chart, token economics draws it in memory bandwidth and Big-O, and multi-agent systems draw it in the roles your agents settle into. In each case the useful work is the same — name the hidden structure so you can shape it on purpose instead of discovering it after it has already hardened.

Further Reading

Things I read recently that didn’t get their own post, all public:

• Software Architecture, AI Engineering, and Conway’s Law (Matthias Jung) — A direct companion to this week’s opener: does the org-structure-to-architecture mapping still hold when AI agents become part of the team topology? Worth reading right after the Conway’s Law post.
• Up to 95% token reduction with zero code — Headroom (Charly Wargnier) — A local proxy that compresses logs, JSON, and code before they reach the model — the preprocessing runs on your machine, though the compressed prompt still goes to the provider. The practical flip side of the token-cost post: if input tokens are the lever, this is one way to pull it.
• The $47K ungoverned AI bill and a multi-provider gateway (Markus Bestehorn) — A LiteLLM gateway for unified access, smart routing, and full cost visibility across Bedrock and other providers. Governance as the first architecture decision once tokens start to cost real money.
• Amazon Bedrock AgentCore is GA (Eashan Kaushik) — Managed agent infrastructure: session-isolated microVM environments, a framework-agnostic runtime, and an MCP Gateway with built-in authentication. The hosting layer under the kind of multi-agent squads the robot-soccer post gestures at.
• Architecting Agentic Solutions with Claude (Vladimir Provorov) — Practitioner patterns for structuring multi-agent solutions. A useful next step if the “roles and shared state” framing landed.
• Who Operates Is the Moat (Adrian Cockcroft) — His read of four agent-orchestration startups: they converge on the same governance spine, so the real differentiator is who holds the pager, not agent cleverness. A Conway-flavored take on where agent architecture is commoditizing.

Until Next Sunday

That’s the week: a law that quietly wins architecture reviews, the hardware constraints underneath what a token costs, and a twenty-year reminder that multi-agent coordination is an old problem with new tools. The thread each time is naming the structure you didn’t design so you can shape it. Which hidden structure in your own systems are you still pretending you chose on purpose?

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.

Learn more →

• WeeklyReview
• AI
• AWS
• Architecture
• MultiAgent