ClaudeHow Ramp engineers work with AI agents at every step
At a glance
WHAT IT’S REALLY ABOUT
Ramp scales engineering velocity using AI agents, workflows, and guardrails
- Ramp used the Fable model to tackle hard monolith problems (import cycles and lazy-loading) and to build empirically verifiable CI improvements through shadow testing and production data validation.
- Dynamic workflows act like “max test-time compute,” orchestrating many sub-agents to profile, optimize, and iterate over days—helping reduce CI P50 from ~18 minutes to ~6 minutes.
- Ramp distinguishes loops (repeatable, known steps like PR babysitting and dead-code cleanup) from dynamic workflows (unknown-step tasks like system optimization and long-horizon experimentation).
- The company built internal agent products—Glass for non-technical users and Inspect as a tool-rich “digital coworker”—embedding agents into Slack/web and enabling multiplayer, VM-backed PR sessions.
- Ramp emphasizes least-privilege access, trace-by-trace debugging over aggregate benchmarks, decentralized automation building, and ROI-first cost thinking rather than per-engineer token budgets.
IDEAS WORTH REMEMBERING
5 ideasPick agent projects with measurable outcomes and low blast radius.
Ramp targeted CI/testing code where performance and correctness can be empirically validated, ran changes in shadow mode, and constrained worst-case impact before widening use.
Use dynamic workflows for problems where the next step is unknown.
They treat dynamic workflows as a higher “thinking dial,” letting the system spawn and orchestrate sub-agents to discover and validate successive optimizations rather than following a fixed plan.
Automations compound when agents can schedule follow-ups and learn from production data.
Fable landed code, scheduled itself to re-run after a day, profiled with real data, and repeated over multiple days until it achieved sustained CI wins and reported results.
Separate “horizontal” repetitive tasks from “vertical” end-to-end ownership.
Loops/routines handle repeated chores (rebasing, fixing CI, code review, dead-code deletion), while agents like Claude Tag/Inspect can run an entire experiment lifecycle with minimal human touchpoints.
Trace-level debugging beats aggregate benchmarks when operationalizing agents.
Ramp focuses on the “correct trace” (which tools/queries/commands should happen) and then adjusts prompts, tools, and access until the agent reliably follows that path.
WORDS WORTH SAVING
5 quotesI told Fable to fix all our import cycles. I also told it to make our app lazy, so the app boots up, and it's an enormous amount of Python modules. And Fable did-- made a lot of progress in both of these. A lot of this code was merged.
— Rahul Sengottuvelu
Just yesterday, it actually reduced our, uh, CI time from, I think, eighteen-minute P fifty to six-minute P fifty.
— Boris
It'll crank up the exposure, make sure the experiment's running, and maybe a couple weeks later, it'll be like, "All right, I'm gonna ship this variant," and it puts up another PR for that, and I wasn't in the loop at all.
— Boris
Especially with the, the latest models, I, I think they, they require much less hand-holding, and sometimes you just gotta let them cook and get out of the way.
— Rahul Sengottuvelu
We've tried to give Inspect all the tools that a Ramp, uh, builder, so product engineering design person would have. So this includes access to GitHub and Linear and Slack and Datadog and Sentry and various other tools.
— Rahul Sengottuvelu
High quality AI-generated summary created from speaker-labeled transcript.
