The Pragmatic Engineer devotes a deepdive to Inspect, the in-house coding agent of fintech Ramp, based on interviews with CTO Rahul Sengottuvelu, head of engineering Hamid Dadkhah, and founding engineer Zach Bruggeman. The authors note that a few companies write most of their code with their own agent: Goose at Block, Minions at Stripe, River at Shopify. They ask: why build when Codex, Claude Code, Cursor, or Copilot already exist?
Inspect is a background agent opened internally in November 2025. Engineers remain free to choose their own tools, but 75% of merged PRs now come from Inspect sessions, up from about 60% in January. The tool stands out on two points. It runs in remote sandboxes, which removes the limit on parallel sessions, centralizes configuration, and enables collaboration. It has the same integrations as a Ramp engineer, so that only the model's intelligence bounds the agent. It verifies its own changes: tests, telemetry, and feature flags on the backend, screenshots and previews on the frontend — a capability built before third-party tools offered it.
Three reasons motivated building it in-house: local machines only allowed one or two Claude Code sessions, the web team wanted frontend tooling for designers, and growing complexity called for remote development environments. The v1, a Chrome extension where you designated a page area to modify, was not adopted: engineers already knew how to edit a file, and the local setup put off non-developers. v2 pivoted to a remote environment with an agent layered on top, relying on OpenCode, an open-source harness exposing an HTTP API and agnostic to the model. Adoption then rose sharply.
Current figures: under 5 seconds to start a provisioned environment, a team of 5.5 people, more than 150 internal contributors, one million sessions passed in July, and more than 80% of Inspect written in Inspect sessions.
Use cases cover small and medium coding tasks, fixing bugs from a Slack thread, and debugging with queries against a sanitized production replica and against Snowflake. More than 200 internal agents run on the platform: ReviewBuddy for code review, Oncall Assistant for incidents, Testo to generate Playwright tests, Ramp Research as a data analyst, Voice of the Customer, and draft PRs created from Sentry or Datadog alerts.
The architectural principle: give agents the same context and the same tools as engineers. The rest, paywalled, details the stack (Cloudflare Durable Objects, Agents SDK, Modal) and the sandbox.