Three numbers, from one fintech company.
As of May, 75% of every pull request merged at Ramp was raised by an agent. 5.5 is the size of the team that built and maintains it: four engineers, a director, and a part-time PM. More than 80% of the agent's own codebase is now written in Inspect sessions.
The tool is called Inspect. Ramp built it in-house and kept building it while Claude Code, Cursor and Codex were all available for the price of a corporate card.
The obvious read is that the frontier tools were not smart enough. The reasons the founding team actually gives say otherwise. Gergely Orosz reported the story with Jessica Salmon and Ivan Klaric. They got CTO Rahul Sengottuvelu, head of engineering Hamid Dadkhah and founding engineer Zach Bruggeman on the record about why Ramp went its own way. Ramp gives three reasons, and not one of them is about how clever the model is. Two are particular to Ramp: the web team wanted better frontend tooling so designers could make small UI tweaks, and growing system complexity created demand for remote dev environments. The third applies to everybody.
Here is what actually caps an agent.
Ramp liked Claude Code on day one. The problem was that a laptop runs one or two agent sessions before it falls over. That is not a model limitation. It is a machine limitation, and it caps the amount of work you can have in flight regardless of how clever the model is.
The second constraint is not on Ramp's list. It is what the reporting shows about most third-party harnesses, and it is the one worth stealing, so take it now, free:
A third-party agent cannot see your telemetry, your feature flags, or your read-only prod replica. It can write the change. It cannot check whether the change worked. So a human does that part, every time, forever — and that human is your actual bottleneck, not the model's reasoning.
Inspect closes that loop. It runs the tests. It reviews telemetry. It queries feature flags. On frontend work, it screenshots its own output and hands back a live preview. Ramp built screenshot verification almost a year ago, months ahead of nearly all the third-party harnesses. They managed it because they were the only ones who could reach their own systems, not because they were ahead on AI.
That is the whole thesis: the ceiling on your agents is access, not intelligence.
The version that failed first
This part tends to disappear from the retelling, even though it is the most useful.
Inspect v1 was a Chrome extension. A designer could highlight a region of the site and tell an AI to make a small change: copy edits or minor UI tweaks. Frontend engineers built it to automate themselves out of the "can you just move this button" loop.
People liked it. Almost nobody used it.
There were two mundane reasons. Engineers already knew how to open a file and edit one line, so the extension saved them nothing. It also required a local development environment, making it too complicated for the non-engineers it was aimed at.
A tool that the experts do not need and the novices cannot set up has no users. That is not an AI problem. It is a product problem, and it killed v1.
So they pivoted. They did not choose a better model. They chose a different substrate. v2, released November 2025, is a remote development environment with a coding agent on top. Central configuration removed the local-setup requirement that had blocked every non-engineer. Remote execution removed the one-or-two-sessions ceiling that had blocked every engineer.
One change fixed both failure modes. Both came from the machine, not the model.
What happened next
- Nov 2025 — v2 ships internally
- Jan 2026 — ~60% of Ramp PRs authored by Inspect
- May 2026 — 75%
- Jul 2026 — 1,000,000 total sessions
Alongside that, ~90% of PRs merged into the Inspect repo itself come from an Inspect session. 150+ Ramp engineers have contributed to the codebase. 200+ internal agents now run on the platform, handling things like code review and incident management. Inspect itself debugs, too: it queries the sanitized read-only prod replica and Snowflake to find business-logic and data mismatches on its own.
And here is the number that makes the rest possible: under five seconds to spin up a fully provisioned remote environment. This is no bare container. Postgres, Redis, RabbitMQ, Temporal, Chromium and VS Code Server are ready.
Five seconds is a product decision disguised as an infrastructure metric. At five minutes, nobody starts a session speculatively. At five seconds, you start ten.
For anyone costing this out, the stack is React/Vite on the front, Cloudflare Durable Objects and SQLite for state, the Cloudflare Agents SDK, Modal for sandboxes, and OpenCode running inside the sandbox. They did not write the agent loop from scratch. They wrote the environment around it.
The cultural detail nobody would have designed on purpose
Every Inspect session is public. Any engineer can open any other engineer's session and collaborate in it.
There is no opt-out.
That is a real decision with a real cost. It is worth sitting with before you copy it. Your half-finished, badly-prompted, going-nowhere agent session is visible to the entire engineering org. Most companies would not survive proposing it.
It is also almost certainly why more than 150 people ended up contributing to a tool owned by five and a half. You cannot learn to drive an agent well from documentation. You learn by watching someone better than you do it, which only happens when sessions are open by default.
What to actually do with this
Do not build your own harness. That is the wrong lesson, and for almost every reader it would be a catastrophic use of a quarter.
Do this instead. Write down the three things you check by hand after every agent run. Write what you actually do, not what you should check. Something like: I run the test suite. I open staging and look at it. I grep the logs for the error I expect.
Now ask whether your agent can reach each one.
That list is the real roadmap. It is usually much shorter and much cheaper than it looks. A test suite the agent can invoke. A staging URL it can screenshot. A read-only log query. None of that requires building Inspect. Most of it is an MCP server and an afternoon.
Ramp is not alone. Stripe built Minions, Block built Goose, and Shopify built River — each independently, each in-house. None of that stays proprietary: the commercial harnesses will close this gap.
So the advantage here has a shelf life. The thing worth having is not the harness. It is the habit of asking what your agent cannot see.
Full reporting: Why Ramp built its own in-house coding agent, Inspect.
