Ramp's own coding agent writes 75% of their merged PRs, maintained by five and a half people

The Pragmatic Engineer published a deep read on August 25 of Inspect, the coding agent Ramp built instead of standardizing on Claude Code or Cursor. The numbers: Inspect authors 75% of Ramp's merged PRs, 90% of PRs in its own repository, passed 1 million total sessions in July 2026, and is maintained by about 5.5 people. Four engineers, a director, and a part-time PM. More than 150 Ramp engineers have contributed code to it.

The stack is named end to end, which is what makes this useful rather than inspirational. React and Vite on Cloudflare Durable Objects with SQLite for state. Modal sandboxes provisioning in under five seconds. The open-source OpenCode harness underneath. They didn't write an agent loop from scratch, they took a commoditized one and spent their five and a half people on everything around it.

That ratio is the thing I keep chewing on. A 5.5-person team supporting 150+ engineers, where the 150 contribute back into the tool. In the old framing that's an internal platform team with terrible leverage. In this framing it's the highest-leverage headcount in the company, and it only works because the loop itself is off the shelf.

Which connects directly to a paper that went up the same day. Researchers read LangChain's deepagents, Earendil's pi and DeepSeek's dsh at pinned commits and followed their commit histories (arXiv 2608.23953). deepagents has been subtracting authored scaffolding over time. pi has been accreting durable infrastructure. Opposite directions, and both arrived at the same five parts: a commoditized loop, an append-only replayable session record, model quirks stored as data, progressive disclosure of context, and explicit extension seams. A held-out third harness shows all five, and in one seam reuses another harness's implementation verbatim, so the authors split their convergence claim into parallel discovery, diffusion, and literal copying rather than pretending it was independent invention. Honest paper.

Ramp's build sits exactly on that shape. The loop is OpenCode. The durable record is Durable Objects plus SQLite. The extension seam is how 150 engineers contribute without owning the core.

The gap the paper names is one nobody has filled: external verifiability. No harness has a tamper-evident record an outside party can check without trusting the runtime. For a company running its own agent that's an internal trust question. The moment you're selling agent output to a customer, or defending it in an audit, it becomes a product requirement.

Paul Dix published "The end of programming" on August 25 arguing the human line-by-line review model is already extinct at the frontier, citing Bun 1.4: multiple agents on a pre-release Fable 5 producing 6,778 commits and over a million lines of Rust in 11 days for about $165,000 in tokens, driven by Jarred Sumner building the parallel-agent scaffolding. He follows up on Simon Willison's site saying that code is now running reliably on millions of developer machines. One practitioner describing his own shipped work, so weigh it as that. The counter he doesn't address is in the same diff: 13,044 unsafe Rust blocks against 73 in comparable hand-written Rust.

Ramp is the version of this argument with receipts and a maintenance budget attached. Dix is the version without one.