At Ramp, 75% of merged PRs come from a harness no vendor sold it

What Ramp built is a development machine with credentials attached. The agent doing the typing inside the sandbox is OpenCode, which Ramp did not write, sitting alongside Postgres, Redis, RabbitMQ, Temporal, Chromium and a VS Code Server ^[7]. The differences Ramp cites against Claude Code or Cursor are the remote sandbox and the internal integrations, with model intelligence named as the only remaining limit on what agents can do ^[5]. The buy-versus-build question, then, is not whether to write an agent. It is whether to own the box the agent runs in and the access that box carries ^[14].

That box has an unglamorous property: it starts in five seconds or less ^[7]. Near-free session startup is the mechanic under running many agents at once, which is the request local laptops could not serve.

The adoption figure deserves reading carefully. Using Inspect is optional, so 75% is a preference signal rather than a mandate, and a quarter of merged pull requests still arrive some other way ^[16]. That residual is the honest scale of the win.

The timing claim is where the strategy gets fragile. Ramp says it shipped screenshot verification almost a year ago, before third-party vendors supported it, and puts itself months ahead of nearly all harnesses ^[18]. Months is the unit. Each capability Ramp built that vendors later shipped stops being a reason to have built it and becomes something to keep maintaining. What does not commoditise is the integration surface, because one company's telemetry and feature flag systems are not a product anyone can sell.

The scope crept the way internal tools do. Inspect began as a Chrome extension for designers prompting minor website changes ^[12] and now covers bugfixing in Slack, debugging, and internal agents for code review and incident management built on the same platform ^[13]. Every session is public and open to collaboration, with no opt-outs permitted [c11b]. That is a rule an internal tool can impose on its own engineers and a vendor agreement generally cannot.

All of this is one company's account, given to the Pragmatic Engineer newsletter by CTO Rahul Sengottuvelu, head of engineering Hamid Dadkhah, and Zach Bruggeman, the founding engineer on Inspect ^[11]. No build cost, dedicated headcount, revert rate or defect rate is published next to the 75%. The named stack is React/Vite, Cloudflare Durable Objects, SQLite, the Cloudflare Agents SDK and Modal sandboxes ^[8], which sketches the shape of the bill without giving its size.

For anyone smaller, the case here rests on two preconditions: internal systems worth wiring an agent into, and enough engineers to keep that wiring current. Without both, an off-the-shelf harness remains the cheaper answer, and the live question is how long the access gap stays open.