The Ramp Playbook — 8 Rules a $44B Company Uses to Let AI Write Half Its Code, Proven on the Cheap Model
The companion playbook to the video. On Aug 6, 2026 Boris Cherny — the creator of Claude Code — sat down with the two people who run Ramp's agent machine, Austin Ray and Rahul Sengottuvelu, and walked through exactly how a $44B company (Series F, June 2026; over $1B in annualized revenue) got to the point where roughly half of all merged pull requests across its frontend and backend repos are started by its in-house agent, Inspect — an agent that is itself now over 80% written by Inspect (per the Modal engineering blog). We pulled the conversation apart into 8 operating rules, then proved the stealable ones live on a real repo using Claude Sonnet, the cheap model. INSIDE: (1) THE MACHINE — Inspect with a human engineer's toolbelt (GitHub, Slack, Linear, Datadog, Sentry, a VM per PR, multiplayer sessions), the dynamic workflow that took CI from an 18-minute P50 to 6 minutes and scheduled ITSELF to re-verify on production data the next day, the AI on-call assistant on Claude Code running on every incident since late February, and the line that frames everything: more agent sessions now come from automations than from humans. (2) ALL 8 RULES, each with verbatim-quote receipts and a 'Steal it' action line — build for the model that's 3-6 months out (scaffolding is tech debt with a countdown timer), treat the agent like a coworker, keep prompts declarative (say WHAT, not HOW — dictating steps caps the agent at your own ability), loops for repetitive work vs dynamic workflows for unpredictable work, least privilege buys agency (read-only service keys, the security team builds the rails AND uses the agents), debug the trace not the benchmark (find the command it should have run, fix one thing), start where failure is cheap and success is measurable (CI first, weeks in shadow mode), and never cap the intelligence (no token budgets — positive ROI means minimizing cost is the wrong objective; cut cost with batch APIs and cheaper models for automations instead). (3) THE 3 DEMOS with the exact prompts we ran on Sonnet — the declarative import-cycles fix with a verifiable done-condition, the make-my-test-suite-faster dynamic workflow, and the dead-code-deletion loop — each with what happened on our machine (4 cycles fixed in 6m59s with all 37 tests green; the suite from 82s down to ~2.8s — ~27x — after 3 fixes raced in parallel worktrees; 2 dead-code commits made unattended before the loop scheduled its own next pass), plus the one-evening on-ramp: one loop, one workflow, one guardrail. (4) THE 7-POINT GUARDRAIL CHECKLIST to pin next to your terminal, from read-only keys to trace reviews. Every quote from the source interview on the Claude YouTube channel (Aug 6, 2026); Inspect figures from the Modal engineering blog. Independent playbook from Hyperautomation Labs — not affiliated with Ramp or Anthropic.
Free. No spam. Unsubscribe anytime.