Technology and industry
The stack underneath: compute, platforms, and the companies building them.
17 essays
The Agent Is Not the MoatThe agent gets the product name and the animated workflow. The durable advantage is accumulating underneath it: distinctive signal, legitimate customer access, and a learning loop that makes the next response more relevant.
When Orchestration Ships as a Feature: The Distribution Stack Is Now the MoatOpenAI’s GPT-5.6 Ultra treats multi-agent coordination as a shipped feature. That shifts the bottleneck from model quality to the distribution stack: orchestration, partner routing, and fresh data. If you cannot explain the route from buyer intent to completed transaction, you do not have a moat. Y…
The Storefront That Closes: HP and Salesforce Are Building the Operating Layer MoatA customer chats. The assistant sounds fluent. Then nothing closes. HP and Salesforce are shipping the missing layer: reusable operating surfaces that connect conversation to real business state. That is the moat shift happening now.
The Distribution Default: Why the Interface That Finishes the Job Beats the Model That Answers FastestWhen model quality drifts toward table stakes, the competitive moat moves to the customer-facing interface that captures intent and turns it into proprietary context. Microagi’s free cleanings for recorded data are the robotics version of what marketing agents must build next.
The Micro-Habit Contract: What Gemini Spark on Mac and Vercel eve Are Really Teaching Desktop AgentsA chat window is a place you visit. A desktop agent with file access is a place work happens. The frontier shift is not model quality. It is the engineering of daily, bounded work units that produce named artifacts in the tools where decisions live.
The Signature Hasn't Moved: Why Duty-Persistence Is the Real Agent Design PatternAI can draft faster than any human review cycle. Courts are now clear that the human duty does not move. That creates a new product design pattern for agentic workflows: preserve state, custody, and provenance from generation to signature.
Design for Wobble: Compute Reliability Is Becoming a Customer Experience SurfaceA France heatwave warmed nuclear reactors enough that operators throttled output. RTE said the grid was secure. Both statements were true, and the contradiction felt like a preview: “always-on” AI agent promises lean on assumptions that capacity can stop guaranteeing. Marketing and customer-facing…
Token Cost Is the Real Infrastructure. The Rest Is Real Estate.Most boards think AI infrastructure means data centers and GPUs. But the real gate is the unit economics of the agentic execution loop: tokens consumed per outcome, retries, tool calls, and re-plans. Caching turns tokens into margin, and digital trade agreements can function like a cost discount fo…
The Agentic Command Layer: Why the Next OS Battle Isn't a PhoneThe next operating system battle will not be a better phone. It will be about who routes your intent into action. Apple, Google, and OpenAI are converging on the same prize: an agentic command layer that executes across the apps you already use. The strategic pressure is now clear, and the leadersh…
Waymo Just Turned a Fleet Liability Into Booked Power for AI ComputeWaymo just treated retired robotaxi batteries like booked capacity for the grid. Not a sustainability story. A procurement-grade reliability layer with delivery timelines. This changes how power gets marketed, underwritten, and planned for AI-adjacent compute.
The Stop Contract: Why the Pause Layer Is Becoming the Real Agentic MoatZigment, Waldo, and IrisAgent launched into adjacent customer-facing categories this week. The market story is agentic CX. The real product story is the stop contract: a legible, editable boundary that decides when the system should act and when it must route to a human. That decision boundary is b…
Compute Geography Is Becoming the Routing Layer for Agentic WorkloadsCompute geography is quietly becoming the routing layer for agentic workloads. When water, power, and interconnect lead times dominate, “siting” stops being procurement trivia and becomes a reliability and token-economics loop.
The Verification Stack Is the New Engineering PlatformThe question is no longer whether your agents can generate artifacts quickly. It is what operational system decides those artifacts are correct enough to ship. Cisco and AWS both point to the same shift: verification is becoming the platform layer for agentic software.
When Agent Deployment Becomes a ZIP File, the Game Moves UpstairsBlueprint skills are becoming portable, runtimes are becoming ZIP-file endpoints, and quality loops are surfacing in platform layers. When agent infrastructure stops being a craft bottleneck, the scarce resource moves upward: evidence of correct outcomes in your environment.
The Agent Moat Is ProofFor two years the AI question was “who has the best model?” Agents change the ordering of priorities. Once agents edit code, touch infrastructure, spend budgets, and trigger workflows, the new operational question is whether you can prove what they did, why they did it, and whether they were allowe…
Owning the Model Is Not Owning the SystemA local Qwen MTP GGUF build hit a runtime assert that looks minor until you notice what it reveals: possession is not control. In AI, autonomy only exists when the system can recover acceptable output after a dependency breaks.


