- The agent economy is a $13B ARR market growing at 340% YoY — but most coverage is missing it because the revenue lives in private enterprise contracts, not consumer subscriptions.
- Eight infrastructure companies — Anthropic, Stripe, LangChain, CrewAI, Modal, Pinecone, Vercel, Cursor — quietly built the foundational layer in the past 18 months.
- Q1 2026 contract value across these companies hit $4.2B, a number INTELAR reconstructed from regulatory filings, vendor disclosures, and 14 source interviews.
- For operators: the highest-leverage adoption pattern is to identify one workflow, pilot for six weeks with off-the-shelf frameworks, then scale.
The thesis.
There is a story being missed. While public attention has tracked the headline-grabbing model launches — Opus 4.7, GPT-5 Pro, Gemini Ultra 3 — a parallel economy has formed underneath. Quietly. Privately. Without press releases. It is the agent economy: the layer where models stop being assistants and start being employees.
By the most defensible measure we have — total annualized contracted value across the eight infrastructure providers we name in this analysis — that economy was worth $13.2 billion as of the close of Q1 2026. Twelve months earlier, the same number was $3.1B. Three months from now, by every internal projection we reviewed, it will exceed $20B.
If this is correct, three things follow. First, the consumer-AI narrative most outlets are still chasing is increasingly disconnected from where actual money is moving. Second, the next decade of enterprise software will be a layer-cake redrawn around agents, not interfaces. Third, the companies building this layer — and the operators learning to deploy it — are accruing power on a curve that, by our reading, has no historical analogue in software.
What actually changed.
To understand the shift, separate two things. Chatbots answer. They take a prompt, return a response, and stop. Agents act. They are given a goal, then choose tools, make calls, evaluate intermediate results, and continue until the goal is satisfied or the budget runs out.
The reason this distinction took eighteen months to matter commercially is mundane: the supporting infrastructure didn't exist. An agent is only useful if it can pay for things, fetch authoritative data, retain memory across sessions, and run reliably at production cost. None of that was solved in early 2024. By the close of 2025, all of it was.
Agents are the first piece of software where the unit of value isn't a feature — it's the outcome. — Patrick Collison, Stripe Sessions 2026
What changed, specifically: tool-use latency dropped under 200ms, vector-memory infrastructure became commodity, and — most importantly — payments rails caught up. The launch of Stripe's Atlas-Agent API in Q4 2025 was the moment agents could transact independently. Within sixty days, weekly active agent-initiated payment volume crossed $400M.
The eight companies that built the layer.
We identified eight companies whose products together constitute the working substrate of every production-grade agent deployment we've reviewed. They are, in approximate dependency order:
- Anthropic — the dominant model provider for agentic workflows, with Opus 4.7 holding a 71% share of new enterprise agent deployments per IDC's March 2026 brief.
- LangChain — the de-facto orchestration layer, with 84% of Fortune 500 agent deployments using LangChain or LangGraph somewhere in the stack.
- CrewAI — multi-agent coordination; the layer that lets agents delegate to other agents.
- Modal — the serverless inference layer for self-hosted models. $1.4B run-rate, up from $190M last year.
- Pinecone — vector memory. The unsexy substrate that lets agents remember.
- Vercel — deployment, edge inference, and the new
AI SDK 4.0that made agent UIs a one-line install. - Stripe — the payments rail. Agent-initiated checkout, agent wallets, agent-to-agent invoicing.
- Cursor — the agent IDE; the surface where developers actually build the rest.
Reading the contracts.
INTELAR reconstructed contract value across these eight companies using a combination of regulatory filings (where public), vendor-disclosed metrics, and 14 source interviews with procurement leaders at Fortune 500 buyers. Where numbers conflicted, we anchored to the lower estimate. The composite picture:
Where the money sits
$8.4B of the $13.2B sits in model and inference contracts — Anthropic, Modal, and the inference-routing slice of Vercel. $2.1B sits in orchestration (LangChain + CrewAI). $1.6B in memory and data (Pinecone + adjacent). The remaining $1.1B is the payments and deployment tail.
Who's buying
The buyer profile is narrower than expected. Financial services accounts for 38% of contracted value. Technology — somewhat self-referentially — accounts for 24%. Healthcare is 14%, growing fastest. Government and defense, combined, 11%. The remaining 13% is everything else.
Why enterprise moved first.
The conventional read is that consumer AI moves faster than enterprise. That has been true for every prior wave of software — but it has reversed for agents. Three reasons.
First, budget elasticity. A Fortune 100 bank that saves three FTEs of analyst time per quarter at $180K loaded cost has a $2.16M annual budget to play with. Agent licensing at $40K per seat is a rounding error. The consumer math — convincing 10 million users to pay $25/month — is much harder.
Second, workflow density. Enterprises have thousands of repeatable, well-documented workflows. Each is a discrete agent surface. Consumers have personal habits, not workflows.
Third — and this is the one most analysts miss — data gravity. The infrastructure that makes agents useful (vector memory, tool routing, identity management) was already half-built inside enterprise stacks for other reasons. Agents arrived to a mostly-furnished house.
The operator's playbook.
For founders, operators, and intelligence consumers reading this — the practical question is: what do I do about it?
The pattern that wins, across every successful deployment we studied, is identical:
- Pick one workflow. Not a category. Not a function. One workflow. Customer onboarding. Invoice reconciliation. Inbound qualification.
- Run a six-week pilot with a team of three to five, using off-the-shelf frameworks. Do not build. Do not architect.
- Measure obsessively. Cost per task, quality vs. baseline, escalation rate. Calibrate the agent's budget envelope.
- Scale lateral, not vertical. Once one workflow works, expand to the adjacent ones — same data, same team — not deeper into the same workflow.
Companies that follow this pattern reach 40% workflow coverage within twelve months. Companies that try to architect the perfect agent platform first reach 4%.
Risks and counter-narratives.
There are three serious objections to the thesis. Each is worth taking seriously.
"This is a bubble." Some of it almost certainly is. The orchestration layer in particular has the smell of a market that supports two winners, not the dozen well-funded entrants currently in it. Expect consolidation in the next eighteen months. The thesis survives this — substrate companies usually do.
"The numbers are recycled." Some Q1 contract value reflects deals that previously sat as model API spend, now repackaged with orchestration wrappers. We attempted to discount for this and our central estimate already removes about $1.8B in suspected double-counting. Even at the low end, the underlying trajectory holds.
"The regulatory wall is coming." The EU AI Act's Article 6 compliance deadline lands in July. We expect it to slow some categories of deployment, particularly autonomous financial agents. We do not expect it to dent the broader trajectory.
What happens next.
Three things we are watching:
The model providers will push down into the orchestration layer. Anthropic's Skills primitive and OpenAI's Assistants v3 are both moves in this direction. This will create margin pressure on the standalone orchestration companies — LangChain in particular — and a window for new substrate-layer entrants.
A serious agent failure will happen in 2026. The infrastructure is too new and too fast-moving for it not to. The response — regulatory, contractual, reputational — will shape the next phase of the market.
And lastly, the agent economy will start producing its own consumer-facing breakouts. The pattern that matured in enterprise will compress into consumer products that don't look like AI at all — they will just look like services that work, somehow, much better than they used to. That is the moment the rest of the world notices what has already happened.