the pyramid you feel is real but it's not the one they're selling you
here is the shape of the thing. not a pyramid exactly, more like a funnel turned upside down with a bottleneck at the top where all the money pools and nobody builds anything useful. you can feel it in every pitch deck, every funding announcement, every breathless post about autonomous agents that will change everything. something is wrong with the economics.
fifteen months ago i would have called it a pyramid scheme and moved on. simple. satisfying. wrong.
the truth is weirder and more useful. the middle layer is pyramidal and collapsing. the poles are printing money. and the infrastructure underneath everything is quietly taxing the entire attempt at a civilization.
let me walk you through the four frames that explain this mess. by the end you will see the game everyone else is missing.
the wrapper graveyard is already full
85% of startups that raised a million dollars or more for "basic llm integrations" failed, shut down, or pivoted by 2025. not struggling, not pivoting to find product-market fit. dead. the cb insights market map shows an explosion from 300 to thousands of players in eighteen months... and the funding data shows the other side of that curve. seed valuations averaging $17.9 million while series a investors now demand two to three million in actual recurring revenue and quietly track something called "human intervention rate" as the metric that matters.
crewai raised eighteen million dollars. the hacker news thread underneath the announcement was savage in that particular way only a room full of builders can be savage. the top comment just said: "i'm struggling to understand the value beyond what you can do with vanilla api calls."
that one line is the crack in the whole industry.
the shovel sellers are being eaten from below by commoditization and from above by platform bundling. aws just put openai codex on bedrock. no procurement friction. no new vendor to approve. one click and your aws-anchored enterprise has agent capabilities baked into the stack they already pay for. the control point was never in the agent framework. it was always in identity, procurement, and observability. the middle skipped leg day.
the gold IS being dug ... just not where you're looking
this is the part that makes the pyramid narrative incomplete. at two poles, real money is moving.
the coding agents are not hype. coinbase reduced time from idea to production by 90% using cursor. 75% of all pull requests are agent-created. teams of one or two engineers now build features that previously required entire teams. not someday, not in a pilot program, right now. rakuten compressed feature delivery from 24 days to 5, a 79% reduction, with claude code executing a complex refactoring on a 12.5-million-line codebase autonomously at 99.9% accuracy.
on the other pole, vertical agents with regulatory moats are quietly stacking revenue. harvey in legal. nabla in healthcare. glean in enterprise search. these are not horizontal platforms that promise to do everything. they do one thing in an industry where compliance is the moat and generic agents cannot touch hipaa or finra requirements.
the lumpy truth sits in the sales development space. sdr agents... the most-funded vertical... have a median first-year roi of 0.4 to 1.6 times cost versus vendor-claimed 3 to 5 times. the enterprises that built custom or hybrid solutions are winning. the ones that bought generic horizontal agents are losing. the money is at the extremes. deep vertical or deep infrastructure. nothing in the middle survives.
the community is not wrong, it is early
developer skepticism is the diagnostic layer most founders refuse to read. the hn thread "ask hn: is anybody getting value from ai agents?" is a gold mine of specific technical failure modes. agents diverge when put in loops. token costs rack up to thirty dollars a day just testing. most frameworks hit a toy app ceiling and cannot handle asynchronous long-running business state. the most damning signal sits in a single sentence from a builder with nothing to sell: "from what i've seen, ai agents aren't even close to being reliable for real-world tasks yet."
but here is where the story splinters. while hn was calling crewai worthless, sap was deploying 200+ agents across finance and supply chain. td bank was compressing 15 hours of processing into 3 minutes. kpmg was rolling out agents to 276,000 employees.
the gap between what developers on forums say and what fortune 500 cios are doing is a structural opportunity hiding in plain sight. consumer-grade agent products fail. enterprise-grade custom deployments work. the shovel sellers selling to indie developers and startups are losing. the shovel sellers selling to enterprises with procurement contracts are winning. this bifurcation is the whole game.
the infrastructure tax is the real pyramid
this is the ugliest frame and the most important one. in the ai agent economy, the real money flows to infrastructure, not applications. compute, orchestration, security, deployment... these layers capture value even when the applications built on top of them fail.
the first quarter of 2026 was not a quarter for building. it was a quarter for feeding the maw. three hundred billion dollars in global venture funding. two hundred forty-two billion to ai. four frontier labs... openai, anthropic, xai, waymo... collectively raised $188 billion. that is 65% of all global venture investment in the quarter. the anthropic/spacex fifteen-billion-dollar-a-year data center deal was met with "alarm, not admiration" by the builder community.
what makes this a tax rather than healthy investment is the concentration. openai and anthropic together took $217 billion in the first half of 2026. that is 43% of everything raised globally. the application layer is starved. the security and trust layer... verifiable agent identity, agent-to-agent discovery, audit trails, micropayment rails... is in alex chen's words the "most underinvested" area of the entire stack.
the infrastructure providers do not need applications to succeed. they just need more applications to be attempted. every failed startup, every pivot, every abandoned agent framework... the compute was already paid for. the money flows up regardless of whether anything works downstream. that is the pyramid structure. not a deliberate scam, but a gravitational pull of capital toward the bottleneck that makes the bottleneck owners rich while everyone else starves.
what survives when the middle burns
three opportunities are real underneath the carnage. none are solo-founder-friendly without serious domain expertise, but they exist because the fundamental needs exist.
vertical agents with regulatory moats. enterprises in law, medicine, and finance need agents that handle compliance, not just generate plausible text. harvey, nabla, and glean are the only funded series a survivors in their spaces. incumbents like salesforce and epic will build native eventually. the window is 18 to 24 months. the regulation is the moat.
custom enterprise deployment studios. 60 to 70% of large organizations are deploying or piloting agents, but only 16% of those deployments qualify as true agents. most are stuck at fixed-sequence workflows. the gap between "we bought an agent platform" and "it actually works in our environment" is where the money lives. accenture and deloitte are capturing it through services. no product-led studio exists yet.
agent trust infrastructure. every autonomous agent will need verifiable identity, audit trails, and secure discovery. agent-to-agent transactions require micropayment rails. this layer is almost entirely empty. the infrastructure tax has pulled all the money upward. the trust layer has been left behind. someone is going to build it.
the shovel makers buried themselves
the ai agent economy is not a pyramid scheme in the fraud sense. it is a pyramid structure in the economic gravity sense. the infrastructure layer taxes everything. the middle layer sells shovels to each other and dies. the gold is real but concentrated at the poles where measurable enterprise roi or regulatory moats provide actual defensibility.
the tell is in the funding. $188 billion to four companies that provide picks and shovels. a graveyard of thousands of agent startups that burned through seed rounds building wrappers. and a handful of quiet winners at the edges who figured out that the gold is not in selling tools to miners. it is in being the mine.
if you are building in the middle of this market right now... selling general-purpose agent orchestration, or another wrapper around someone else's api, or a horizontal platform that promises to do everything for everyone... you are not a gold miner. you are not even a shovel seller. you are standing in a graveyard holding a shovel you bought from someone else and wondering why nobody showed up to dig.
the gravediggers, at least, are making a killing.