Ramp built its own coding agent with 5.5 people
The fintech turned down off-the-shelf coding agents and wrote its own. It says 75% of merged pull requests now come from it, and the team that maintains it is four engineers, a director and a part-time PM.
Most companies buy their AI coding tools. Ramp wrote one.
The tool is called Inspect. Ramp's engineering leadership described it to the Pragmatic Engineer newsletter as a coding agent that runs on remote sandboxes with access to the company's internal data sources.
The trigger was not ambition. It was a ceiling. Ramp's engineers liked Claude Code on day one, according to the account they gave, but local machines would only run one or two sessions at a time. Engineers wanted more agents running in parallel than a laptop allows.
Two other pressures pushed the same way. The web team wanted better frontend tooling so designers could make small UI tweaks themselves. And as the codebase grew, more work sat at the seams between systems — backward compatibility, broken API contracts — which pointed towards remote development environments rather than local ones.
The first attempt failed. Version one was a Chrome extension that let a user highlight part of a page and describe the change they wanted. Ramp says people liked it and did not use it: engineers already knew how to open a file and edit a line, and the extension still required a local development environment, which put non-developers off.
Version two, released in November 2025, moved the whole thing into a remote environment configured centrally. That removed per-machine setup. The team did not write the agent itself either — Inspect uses OpenCode, the open-source harness, which exposed an HTTP API and was model-agnostic.
The adoption figures are Ramp's own. By January, two months after launch, around 60% of the company's pull requests were authored by Inspect. By May that was 75%. Ramp says roughly 90% of PRs merged into Inspect's own repository come from an Inspect session, and that more than 80% of Inspect is written in Inspect. It passed one million total sessions in July.
The maintenance bill, on Ramp's account, is 5.5 people: four engineers, a director and a part-time product manager. More than 150 engineers have contributed to the codebase, and every session is public with no opt-out.
The defensible part is not the agent. It is the plumbing. Inspect verifies its own backend changes by running tests, reading telemetry and querying feature flags, and its frontend work by producing screenshots and live previews. Third-party tools generally cannot do that, because they are not wired into a company's telemetry or feature flag systems.
Ramp built screenshot verification about a year before third-party vendors supported it, which is where the months-ahead claim comes from.
The pattern is worth noting by anyone weighing the same decision. Ramp did not build a model or a harness; it bought both, then spent its own engineering time on internal integrations nobody else can sell it.
The second-order effect is the one to watch. Engineers have since built more than 200 agents on top of Inspect, including a code review tool one engineer wrote in a week. The build paid for a platform, not a product.