When you map out how companies are actually building their agent infrastructure — from Stripe to Ramp to Spotify to five-person startups — a striking pattern emerges. Almost everything is blue. Almost everything is built in-house.
The harness is commoditized. It is either Claude Code, open-source forks, or the Anthropic Agent SDK. Nobody is differentiating here. Triggers and services are bought off the shelf. LLM gateways are bought. But the middle of the stack — context management, memory, skills, orchestration, and session state — is overwhelmingly homegrown. Company after company is building their own version of these layers because nothing on the market solves them adequately.
The maturity of the company's existing dev tooling predicts the pattern. Mature organizations like Stripe had existing infrastructure they could repurpose — they forked Goose, plugged it into their CI/CD, and were off to the races. Less mature organizations like Ramp pulled nearly everything off the shelf — Modal for sandboxing, Cloudflare for orchestration, GitHub for auth. They look like the most complex implementations precisely because they had to buy every piece rather than reuse what they already had.
The smaller and less mature the company, the more likely they are to buy. The larger and more mature, the more likely they are to build. This creates an interesting strategic opening: if you can give the earthlies of the world — small teams building their first agent — a platform where all of these middle layers are already built, with enough flexibility to adapt, you capture them at the moment they need you most and grow with them as they mature.
But there is a warning embedded in the data. Companies that over-indexed on buying everything off the shelf will eventually realize they should be building and maintaining these layers themselves. If you watch those blue dots and they do not slowly turn green over time, something is wrong. I've argued elsewhere that the organizational context gap is the universal piece every team is forced to build — and it is the most durable place to stand if you can solve it.
Sources
Related Essays
Homegrown Platforms Decay
Internal agent platforms are built by ambitious individuals with other jobs. When those engineers move on, the platform becomes a liability.
The Gap Is Infrastructure, Not Intelligence
The distance between an AI demo and an AI deployment is not a model gap or a harness gap. It is the absence of composable primitives for the boring parts of operationalization.
The Agent Stack Build-vs-Buy Map
Lay out the seven layers of the agent stack and a clear map emerges. The harness is commoditized. Context, memory, and orchestration are blue across the chart.
Key takeaways
- Across companies of every size, the agent harness is bought and the middle of the stack — context, memory, skills, orchestration — is built in-house.
- Mature companies build because they can repurpose existing dev infrastructure. Less mature companies buy because they have to.
- The strategic opening is to give early-stage teams a platform with the middle layers already built, then grow with them as they mature.
FAQ
Why is the middle of the stack built rather than bought?
Because nothing on the market solves it adequately yet. Context management, memory, skills, orchestration, and session state all touch organizational specifics that off-the-shelf tools do not understand. The result is that company after company writes their own version of the same components.
Should mature companies still consider buying these layers?
Yes — eventually. Companies that over-index on building today often discover they are maintaining infrastructure that vendors will out-ship within a year. The right question is which layers are differentiating versus which are merely table stakes that you got stuck with.