Ramp's AI coding agent now merges 75% of its PRs

Ramp's engineering team built its own AI coding agent, Inspect, after concluding that off-the-shelf tools like Claude Code and Cursor couldn't run enough concurrent sessions or work in remote development environments. The fintech company says Inspect now authors 75% of its merged pull requests, up from 60% in January, and has logged more than a million total agent sessions as of July.

Inspect runs on remote sandboxes built with Cloudflare Durable Objects, SQLite, and Modal, giving each session access to the same databases, internal APIs, and browser tooling a human engineer would use, plus screenshot-based visual verification for frontend changes. The underlying agent harness is OpenCode, chosen for being open source and model-agnostic rather than locked to one vendor. New environments spin up in under five seconds.

The core reasoning behind the build-versus-buy call: existing coding agents were capped at one or two local sessions, and Ramp's web team wanted designers to make UI tweaks without waiting on an engineer. More than 150 engineers have since contributed to Inspect, and over 80% of its own codebase was written using Inspect sessions. The platform team, now 5.5 people, has spun off more than 200 additional internal agents built on the same infrastructure, covering code review, incident triage, and customer feedback analysis.

For teams evaluating whether to adopt an existing coding agent or build a custom one, Ramp's numbers are a data point on the concurrency and remote-environment gap in current commercial tools, not a case that every team needs its own Inspect.