Summary
This video from Anthropic’s Claude channel features Ramp engineers discussing how they have integrated AI agents — specifically Fable, a Claude-based coding agent — into their day-to-day engineering workflows at production scale. The conversation goes well beyond surface-level AI adoption, digging into architecture decisions, trust models, and measurable outcomes.
A standout example: Ramp’s team used Fable to optimize their CI pipeline on a large monolithic Python codebase. The agent autonomously profiled performance, landed a series of optimizations over multiple days (scheduling itself to return and check production data), and ultimately reduced CI P50 time from 18 minutes to 6 minutes — a 66% improvement with no direct human intervention in the optimization loop. The engineers also describe using Fable to fix import cycles and lazy-load Python modules at startup.
The engineers articulate a clear mental model distinguishing agent loops (repetitive, well-defined tasks like babysitting pull requests or rebasing) from dynamic workflows (open-ended tasks like system optimization where the steps are not known in advance). They discuss how they implement least-privilege access controls — giving agents read-only service keys to BigQuery and Datadog, for example — and explain their focus on studying individual agent traces rather than aggregate benchmarks to diagnose where agent behavior diverges from the intended path. The result is a detailed, experience-grounded picture of what responsible, high-agency AI deployment looks like at a fast-moving fintech company.
📺 Source: Claude · Published August 06, 2026
🏷️ Format: Workflow Case Study
