Back to YouTube
Parker RexJanuary 12, 2026

Ramp Built Their Own Devin

Ramp open-sourced their internal coding agent 'Inspect' that writes 30% of their PRs. Parker breaks down the architecture, why custom harnesses beat off-the-shelf tools, and what you can learn from their blueprint.

Watch on YouTubeSubscribe to Parker Rex →

Show Notes

Ramp built their own Devin-style coding agent called Inspect that writes 30% of their merged pull requests. They've open-sourced the full blueprint—and since they're not an AI company, they have no reason to hide their secrets.

Why Ramp built their own agent

  • Off-the-shelf tools will never match the context and tooling specific to your codebase.
  • Custom harnesses can have deep hooks into your stack (Sentry, DataDog, LaunchDarkly, GitHub, Slack, Buildkite).
  • When you own the tooling, you build something significantly more powerful than generic solutions.
  • The moat isn't in the LLM—it's in the integration layer.

The results

  • 30% of merged PRs (both frontend and backend) are written by Inspect.
  • Reached this level in just a couple of months.
  • No one was forced to use it—engineers adopted it because it worked.
  • 18 engineers using it concurrently in a shared session view.

Key insight: Agents should have agency

We made sure Inspect is never limited by missing context or tools, but only by model intelligence itself.

This is a bold take. They give the agent everything a Ramp engineer would have—same dev environment, same tools, same access.

Architecture highlights

Near-instant startup

  • Repo images rebuild every 30 minutes.
  • Sessions spin up from stored snapshots.
  • Optimistic warm-up: starts cloning and setting up before the user finishes typing.

Smart sync strategy

  • In large repos, incoming prompts rarely touch files changed in the last 30 minutes.
  • Agent can start reading files immediately while sync completes.
  • Write/edit calls are blocked until sync finishes.

Tech stack

  • OpenCode — Server-first typed SDK with plugin system for the agent loop
  • Modal — Sandbox VMs with everything an engineer would have locally
  • Cloudflare Durable Objects — Per-session SQLite databases for high performance
  • Cloudflare Workers — WebSocket hibernation API for real-time streaming without idle compute costs

Surfaces

  • Slack
  • Chrome extension
  • VS Code
  • PR comments

Why custom beats off-the-shelf

  • It only has to work on your code.
  • Deep integration with your observability stack (Sentry, DataDog).
  • Encodes how your team ships—not generic best practices.
  • No waiting for vendors to add features you need.

The developer-as-operator shift

  • Developers are now agent operators.
  • Better leverage on agents means better outcomes.
  • Time-to-productivity for new engineers improves on both sides: better human onboarding docs and better agent instructions.
  • Multiple agents can work on non-overlapping parts of the codebase concurrently.

Should you build your own?

The classic engineering answer: it depends.

  • If you're not at Ramp's scale, off-the-shelf might be fine.
  • But serious engineering teams will likely end up building custom harnesses.
  • You're building on the shoulders of giants—they've shared the blueprint.

What this means for 2026

2026 is actually the year of the background agent.

The proofs are in the code. When 30% of PRs at a company like Ramp are written by an agent, it's not hype anymore.

Links

Next showPrevious show