I just spent time going deep on how real companies — Ramp, Stripe, Spotify, Coinbase, Earthly — are actually building their agent infrastructure. Not what they say on stage. What they actually deployed. The picture is striking. The agent stack is almost entirely homegrown. There is very little purchasing happening at the layers that matter.
Lay out the seven layers — harness, tools and file ops, context and memory, orchestration and session state, sandbox and execution, LLM gateway, observability and evals — and a clear map emerges.
The harness is commoditized. With few exceptions, every company is on Claude Code or OpenCode. Spotify migrated from a homegrown loop to Goose to Aider to Claude Code. Bitrise hand-rolled because they are literally a CI/CD platform. Everyone else converged. This layer is done. There is no company to build here. Even the players who raised money to compete are fighting over a slot that is becoming a commodity feature of existing platforms. The harness is a force multiplier, not a standalone business.
The LLM gateway is bought. Triggers and external integrations are mostly bought. Observability is sometimes bought — Coinbase uses LangSmith — but more often built. None of these layers is where the durable category sits.
Context, memory, skills, orchestration, and session state? Almost entirely built in-house. Across every company I studied, regardless of size, these layers are blue on the chart. Nobody has a good off-the-shelf answer for how agents accumulate organizational context. Nobody has a good off-the-shelf answer for how agents coordinate multi-step work. That is the real gap. Not the harness. Not the integrations. The connective tissue that makes agents operationally useful — and the layer where I've argued organizational context is the hardest unsolved problem.
If you are picking what to build, the build-vs-buy map is the most useful planning tool you can draw. Buy where the category is settled. Build where every serious operator is forced to build. Skip the middle.
Sources
Related Essays
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.
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.
The Real Product Is What Replaces Homegrown
Do not compete with what Shopify is building internally. Build the system that replaces every homegrown agent platform when the engineer who built it moves on.
Key takeaways
- The harness layer is commoditized. Claude Code or OpenCode at almost every serious team.
- LLM gateway, triggers, external integrations are mostly bought.
- Context, memory, orchestration, and session state are built in-house almost universally. That is the gap.
FAQ
Why is the harness layer commoditized?
Open source and incumbent vendors converged on a small number of options that are good enough that nobody is gaining a sustainable edge by writing their own. Spotify, Coinbase, Stripe — they all converged.
Why is everyone building context and orchestration in-house?
Because no vendor has shipped a credible answer for how organizational context flows to agents or how multi-step work coordinates across them. The category does not exist as a buyable product yet.