The through-line this week is convergence at a frequency that feels like consensus: the AI economy is generating spectacular proof of capability and almost no auditable proof of durable revenue. From indie agent builders stacking Mac Minis to OpenAI's 900-million-user engagement problem to Anthropic's distillation accusations, every signal points the same direction — the cost of building v1 has collapsed, but the cost of everything that comes after hasn't moved. We're watching an industry discover, in real time, that the hard part was never the technology.
The Agent Gold Rush Is Long on Screenshots, Short on Receipts
The AI agent hype cycle has reached a recognizable phase: the part where the loudest success stories are indistinguishable from marketing.1 Search volume for "AI agents passive income" is spiking. Tech Twitter is saturated with Mac Mini stacks and OpenClaw dashboards radiating implied wealth. An aesthetic economy has matured around agents — screenshots as product, repos as social proof, dashboard glow as revenue proxy.
The documented cases of durable, agent-driven income remain conspicuously thin. The auto-trading narrative — your agent detects market inefficiencies while you sleep — founders on a basic asymmetry: if the inefficiency is obvious enough for a Mac Mini to find, a quant fund with real infrastructure found it last Tuesday. Widely shared strategies don't stay profitable. They become transfer-of-wealth mechanisms, usually not in the sharer's favor.
Where agents are generating real revenue is exactly where you'd expect and exactly where no one goes viral: back-office automation, compliance documentation, lead qualification, reconciliation workflows. Boring, vertical, sticky. Companies pay for friction removal. They don't pay for aesthetic proximity to alpha.
This is the vibes-vs-value gap hardening into something structural. The pattern hasn't shifted — it's just gotten sharper. The people consistently making money from the agent boom are selling infrastructure, orchestration tooling, and vertical automation. The shovels are fine. The livestreamed gold rush remains unpredictable.
Coding Agents Multiply What You Already Know
Drew Breunig's framing of coding agents cuts through the noise with a pair of observations that should be required reading.
First: skilled developers dramatically underestimate the intuitive knowledge they bring to their prompts. When an experienced engineer says Claude Code "just worked," what you're not seeing is the prompt — relatively specific, using the right terms, activating entirely different model weights than someone typing "the search is broken fix it." The luminaries who say they haven't written code in weeks aren't lying. They're just not seeing how much domain expertise they inject into every interaction. Expertise is invisible to the person who has it.
Second: most of what gets hyped as agentic coding output is personal software, not products. Breunig built an RSS-feed-finding browser extension by forking an existing project and pointing Claude Code at it. He uses it daily. He will not support it, push it to app stores, test it across browsers, or build it for users who don't share his specific setup. "Code is free, as in puppies." The v1 is trivial. Everything that makes it a product — testing, support, cross-platform, marketing — remains expensive.
This is the first-version trap at the individual scale, and it connects directly to the receipts problem above. AI has collapsed the cost of manifestation without proportionally reducing the cost of maintenance, support, or market fit. The result is an explosion of personal tools that look like products but aren't — impressive, useful, and fundamentally non-transferable. The people shipping agent-written code into actual products are testing, reviewing, supporting, and doing all the unglamorous work that "coding is solved" rhetoric conveniently omits.
Opus 4.6 Is Impressive. The Structural Question Is Whether That Matters.
Anthropic's Claude Opus 4.6 landed, and by most accounts it represents a genuine capability jump — particularly in agentic reasoning and sustained multi-step task completion. This isn't disputed.
What matters more is the context it lands in. Benedict Evans's strategic analysis of OpenAI doubles as a diagnosis of the entire foundation model layer: roughly half a dozen organizations are shipping competitive frontier models, leapfrogging each other every few weeks, and there is no known mechanic for any of them to build a durable lead. No network effects. No self-reinforcing market share. No equivalent of what Windows, iOS, or Google Search had.
Evans draws the browser analogy, and it's the sharpest articulation of the commoditization problem to date. A chatbot, like a browser, is an input box and an output box. The last real product innovations in browsers were tabs and merging search with the URL bar. Microsoft won browsers for the first generation of the consumer internet, and it turned out not to matter — the value accrued elsewhere, to the experiences built on top. The same structural pressure is now bearing down on foundation models.
For practitioners, the implication is blunt: build for the capability tier, not the specific model. Whatever you design around Opus 4.6's strengths today will need to survive the field catching up by summer.
OpenAI's Flywheel Is Actually a Treadmill
Evans's full strategic audit deserves its own section because the diagnosis extends beyond one company to the shape of the market.
OpenAI has 800–900 million weekly active users, undifferentiated technology, and shallow engagement. Eighty percent of ChatGPT users sent fewer than 1,000 messages in all of 2025 — an average of less than three prompts a day, and many fewer distinct sessions. Only 5% pay. OpenAI itself calls this a "capability gap" between what models can do and what people do with them — which, as Evans notes, is a diplomatic way of avoiding the words "product-market fit."
If people who know what this is and know how to use it still can't think of something to do with it on an average day, a better model may not be what's missing. It may require entirely new experiences that haven't been invented yet. And if that's the case, there's no structural reason the model provider will be the one who invents them.
The CFO's flywheel diagram — capex drives users, users drive revenue, revenue drives capex — describes a treadmill, not a cycle of increasing returns. A flywheel implies each revolution gets easier. Nothing in OpenAI's metrics suggests that. Meanwhile, Google and Meta are gaining market share through distribution advantages with products that look essentially identical to the typical user. Anthropic tops benchmarks but has near-zero consumer awareness. The competition is shifting to brand and distribution, which is exactly what happens when the underlying product resists differentiation.
Sam Altman's response — everything, all at once, yesterday — reads as acute awareness of the problem. Trading paper for durable strategic position before the window closes. Whether that constitutes strategy or just activity remains to be seen.
Distillation: The Tension That Doesn't Resolve
Anthropic publicly accused DeepSeek, Moonshot AI, and MiniMax of industrial-scale distillation — over 16 million exchanges through roughly 24,000 fraudulent accounts. The headline sounds alarming. The technical reality is more nuanced than the political framing suggests.
DeepSeek's 150,000 exchanges are negligible at training scale — a small team running experiments, likely unknown to the broader organization. The volume came from MiniMax (13 million exchanges) and Moonshot (3.4 million), both focused on agentic reasoning and tool use — precisely the capability where Claude currently leads. At an estimated 150–400 billion tokens across the two larger campaigns, this could meaningfully improve post-training. But quantity is a crude measure of impact. Getting outputs from a teacher model to actually improve a student model is a genuine research problem, not a copy-paste operation.
The structural point is more important than the accounting. Distillation is a shortcut to compute, not a replacement for it. As reinforcement learning becomes central to frontier training, the value of distilled outputs diminishes — RL requires on-policy inference from your own model, and that's where most of the compute cost lives. The RL era makes pure distillation a shrinking advantage even as the political rhetoric around it intensifies.
The tension Anthropic can't resolve is fundamental: you cannot simultaneously offer the world's best agentic model as an API product and object when that access trains competitors. If the models are so precious that distillation is existential, the logical endpoint is restricting them to first-party products — which no lab will do because the API revenue is essential. But the political ratchet only turns one direction. Expect other labs to follow with their own disclosure campaigns. The technical merits will matter less than the utility of these narratives as instruments in AI trade policy.
The AI Layoff Script Hardens Into Convention
Jack Dorsey's Block memo — announcing job cuts framed as AI-driven restructuring — continues a pattern that's been building for weeks.2 Block over-hired, is correcting, and is using AI as the narrative that makes a management failure sound like a strategic pivot. The suggestion that the highest-leverage AI-for-human swap at Block would be replacing Dorsey himself lands as dark comedy, but the serious point stands: "we're replacing headcount with AI" is becoming boilerplate. The enterprise version of the Mac Mini stack. A signal designed to imply a story the evidence doesn't support.
AI as corporate narrative device is following the same trajectory as "digital transformation" a decade ago. The informational value of the phrase in a corporate memo is now approximately zero.
What to Watch
The post-demo transition enters its test phase. Every signal this week reinforces that the field is moving from "can we build it?" to "does anyone durably pay for it?" The next quarter will be defined not by capability announcements but by retention data, revenue durability, and whether OpenAI's inch-deep engagement has a product solution or a structural one. If the engagement numbers don't deepen meaningfully, the browser analogy becomes prophecy.
Vertical automation as quiet acquisition target. The real agent revenue is in workflow automation that never trends — compliance, reconciliation, operational stitching. Watch for acquisition activity from larger players who've noticed that the boring stuff is where the money actually lands.
Foundation model pricing hits the floor. With half a dozen competitive frontier models and no differentiation mechanic, commodity dynamics are accelerating. The question isn't whether prices drop but how fast, and what that does to labs whose entire model depends on API margins. The race to the bottom has a floor somewhere. We haven't found it yet.
Distillation as standard trade-policy instrument. Anthropic's disclosure sets a template. The gap between the technical reality (diminishing returns in an RL-dominated training paradigm) and the political utility (justifying restrictions under a national-security banner) is where the actual policy will get made.
Way Enough is written collaboratively by a human and an AI agent.