A year ago, the AI coding decision was which subscription to expense. Cursor or Copilot, seat count, done. Now a handful of companies, including mine, write most of their code with agents they built themselves, and that turns a tooling preference into a strategic call that is easy to get wrong in both directions.
Ramp’s internal AI coding agent, Inspect, authors more than sixty per cent of the company’s merged pull requests. It runs each session in a sandboxed virtual machine with the full local development stack, and it verifies its own work the way a Ramp engineer would, because it has the same context and tools. Block runs Goose, Stripe runs Minions, Shopify runs River. Wall Street English runs Astra. These are not side projects. They are the primary way code gets written at those companies.
Why they built rather than bought
Ramp has been open about the reasoning. They liked Claude Code immediately, but local machines only let them run one or two sessions at a time, and they wanted many in parallel. Their web engineers wanted better front-end tooling so designers could make small changes directly. As the company grew, more of the hard work sat at the seams between systems, which meant agents needed remote environments and access to internal data sources, with a way to prove a change was correct before a human looked at it.
None of those reasons are about the model. The model is a commodity they rent, and they swap it when a better one ships. The reasons are all about the layer around the model: how the agent gets company context, how it runs the tests, how it stands up an environment, how it checks its own output. That layer is where a company’s specifics live, and it is either something you own or something you rent from a vendor whose roadmap is not yours.
The trap in each direction
Over-building. Most organisations are not Ramp. A coding platform is a real team with a real backlog, forever. If your actual constraint is that developers want a better editor experience, buy them one. Building a harness to solve that is a way to acquire a maintenance burden and call it strategy.
Under-building. If agents are already writing a large share of your code and you have no control over how they retrieve context or verify results, you have outsourced a core capability without deciding to. You will feel it the first time the vendor changes its defaults, its pricing, or its verification behaviour, and your delivery throughput moves for reasons you cannot influence.
A framework for where you sit
Four questions, roughly in order of weight.
- What share of your merged change is already agent-authored? The higher it is, the more the harness is production infrastructure and the less comfortable you should be renting all of it.
- Is your verification story specific? Regulated environments, complex internal systems, and unusual compliance needs are poorly served by generic off-the-shelf checking. If a correct change at your company means something particular, you probably need to own the part that decides.
- Can you run the parallelism you need on the tools you can buy? If the answer is no and that constraint is real, it is the strongest single reason to build.
- Do you have the engineering capacity to maintain a platform indefinitely? Not to launch one. To keep it current against a frontier that moves monthly. If you do not, an internal platform will rot into a liability within a year.
The middle path
Most companies that need more than a subscription do not need a full in-house agent. They need to own the context and verification layer: the skills and playbooks that encode how their systems work, the evaluation harness that defines a correct change, the environment provisioning that lets an agent actually run things. Rent the agent loop and the model on top of that.
In practice that means treating your internal context as a product with an owner. Someone maintains the library of how your services fit together, keeps the test and environment setup current, and updates it as the architecture moves. That work is valuable no matter which agent consumes it, and it does not lock you to a vendor. The agent on top can be Claude Code this year and something else next year without the investment resetting.
That captures most of the value Ramp is getting without committing to a platform team whose job never ends.
Name it before it names itself
This is the part that should reach the executive team. The moment agent-authored code is a material fraction of your output, the harness is production infrastructure, whether or not anyone decided that. It has an owner, an on-call rotation it needs, and a failure mode that looks like your whole engineering org slowing down at once.
You can staff it deliberately, or you can consciously pick the vendor you are trusting with it. What you cannot do for much longer is treat it as a tool someone expensed.
Leave a Reply