A useful lens for understanding the agent platform market is a two-by-two: maturity of a company's dev tooling on one axis, and whether they are reusing existing human dev infrastructure or purpose-building for agents on the other. Plot every public case study I've seen onto that grid and the pattern is unmistakable.
Companies with mature dev tooling — Stripe, Spotify, Coinbase — overwhelmingly build their own agent stack. They already have CI/CD pipelines, sandboxes, observability, and governance. Adding an agent harness on top of what they already own is straightforward. They are not buying platforms. They are assembling them from parts they already have. Most are not addressable as customers and never will be.
Companies without mature dev tooling — or those purpose-building for agents rather than reusing existing infra — buy far more off the shelf. Ramp is the extreme example: dispatcher from Cloudflare, orchestrator from Cloudflare again, sandbox from Modal, LLM gateway off the shelf, auth from GitHub, verification from a vendor. Nearly every component was purchased. They are a sophisticated fintech company, but their dev tooling was not built for this, so they bought. The buy cohort is broader, faster-growing, and has actual budget allocated to the line item.
There is a corollary worth noting. When companies build agents for engineers, they need heavy infrastructure — sandboxes with VNC and Chromium, Modal compute, pre-commit hooks, full CI/CD integration. When the same company builds agents for non-engineering roles, the infrastructure requirements collapse. One company's engineering agent stack was a complex web of purchased and integrated services. Their non-engineering agent stack was Slack channels and Okta. As I've argued in the primitives are the same across roles, the engine is shared but the body is light when you move outside engineering.
The implication for anyone building an agent platform: stop pitching Stripe. They already built it. Pitch the company whose internal tooling is a wiki and a deploy script. They are buying, and they are buying soon.
Sources
Related Essays
What the Build-vs-Buy Data Actually Shows
From Stripe to a five-person startup, the agent stack is mostly blue — built in-house. The harness is bought. The middle of the stack is built. The opportunity sits in turning blue dots green.
Agent Build Versus Buy: Why Engineers Keep Building It Themselves
An engineer can ship a working agent harness in a week and a half. Code is cheap. So what does a paid agent platform offer that a week of engineering does not?
The Mega-Agent Fantasy Is Already Falling Apart
Enterprise AI is not one omniscient agent. The mega-agent fantasy collapses on contact with production, and the failure mode is always the same.
Key takeaways
- Mature dev tooling predicts build behavior: Stripe, Spotify, Coinbase assemble agent stacks from parts they already own.
- Companies without that tooling — or those purpose-building for agents — buy nearly every component off the shelf.
- The buy cohort is the real addressable market for an agent platform. The build cohort was never going to be a customer.
FAQ
What does the build-versus-buy axis actually look like?
Two axes: maturity of internal dev tooling, and whether the company is reusing existing human dev infrastructure or purpose-building for agents. Companies high on both axes build. Companies low on either axis buy.
Why does Ramp buy almost everything?
Ramp pulled their dispatcher and orchestrator from Cloudflare, sandbox from Modal, LLM gateway off the shelf, auth from GitHub, verification from a vendor. They are sophisticated, but their dev tooling was not purpose-built for agents — so they assembled rather than built.