In 2008, if you wanted to run a web application, you leased a server, configured the OS by hand, patched it, load-balanced it yourself, and prayed it survived a traffic spike. By 2015 you clicked a button in a console and Amazon did all of it — provisioning, scaling, healing, metering you by the second. The button was worth more than most of the companies that pressed it. AWS became one of the most profitable businesses on earth not by building the websites everyone visited, but by owning the layer every website ran on.
The agent economy is at its 2008 moment. The button is being built right now — and, just like last time, it is not the thing getting the headlines.
Agents do not run in isolation. A real workload is never one clever model answering one prompt. It is a dozen agents handing work to each other, retrying failures, escalating edge cases, waiting on each other’s output, and recovering when a step dies at three in the morning. Someone has to schedule that, route it, hold the state, and recover it. That someone is orchestration — the first layer of the dependency stack: lifecycle, coordination, and inter-agent communication. The platform that owns it will look like AWS did in 2010: model-agnostic infrastructure, not a model.
Take a workflow that sounds trivial — issuing a customer refund. In production it is not one agent. It is an intake agent that reads the request, a policy agent that checks eligibility, a payments agent that moves the money, a ledger agent that records it, a comms agent that emails the customer, and a fraud agent with veto power over all of them. They run partly in parallel, partly in sequence, and on any given run one of them will fail — a timeout, a rate limit, a malformed response. Multiply that by every workflow in a company and you have a distributed-systems problem, not a prompt problem. Strip orchestration away and every team rebuilds the same plumbing — queues, retries, state machines, hand-offs, observability hooks — from scratch. Nobody wants to write that twice, and work nobody wants to write twice is exactly what gets standardised into a platform.
Cloud computing is the template, almost beat for beat. In 2006 the landscape was fragmented — AWS, Google App Engine, Rackspace, Joyent, GoGrid, all plausible. By 2012 AWS was clearly dominant. By 2018, AWS, Azure and Google controlled roughly 67% of the market: a six-year fragmentation-to-dominance cycle. Zoom out and the ratio is even more telling. Combined revenue from internet infrastructure — CDN, DNS, SSL, cloud — exceeds $300 billion a year today, and infrastructure markets historically take five to seven years to materialise after the application layer emerges, then surpass it. The application layer gets built first because it is visible and fundable; the infrastructure underneath is nascent and unglamorous, right up until it is enormous. Agent orchestration in 2025 looks like cloud in 2008: crowded, early, and one consolidation away from a winner.
Watch the capital, not the demos. LangChain had already raised at a $1.25 billion valuation as of late 2025 — not for a model, for the orchestration layer. CrewAI, AutoGen, Semantic Kernel, Haystack, the Claude Agent SDK, and OpenAI’s Agents API are all racing for the same primitive. My dated call, PRED-004, is that one platform — or a two-to-three player oligopoly — takes 30% or more of the agent-orchestration market by December 2029. And PRED-003 sizes the ground it sits on: the agent dependency layer becomes a $100 billion-plus market by 2030, bigger than the agent apps built on top of it. That is the structural bet — the layer, not the apps.
The tells will be boring and specific. Funding flowing into orchestration and coordination rather than into another wrapper on a model. A platform quietly becoming the default new teams reach for without evaluating alternatives. Standard primitives — a common way to define an agent, hand off a task, checkpoint state — emerging the way HTTP verbs or S3 buckets became defaults nobody argues about. When the primitives standardise, the winner is close.
Everyone is watching the model wars. But the model you use is becoming a runtime detail — swappable, commoditising, cheaper roughly ten times every two years. Betting the company on which model is best this quarter is like betting on which brand of server AWS would rack. The orchestration layer is where lock-in and margin accrue, and it wins precisely by not being a model. Neutrality is the moat: the platform that stays agnostic gets to sit under everyone.
The trillion-dollar businesses of this economy will not own the best model. They will own the layer the models plug into.
The AI Agent Economy — Why the Next Trillion-Dollar Industry Has No Employees is out now on Amazon and Notion Press. This series expands Chapter 2: the dependency layer.
Go deeper: the orchestration prediction behind this arc → https://atin-agarwal.com/predictions/pred-004-orchestration-platform-dominance/?utm_source=substack&utm_medium=newsletter&utm_campaign=aae-16
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.