Quality assurance · Production

Ramp builds Inspect: an internal background coding agent that writes ~30% of merged pull requests

The problem

Ramp needed a coding agent that not only writes code but also closes the loop on verifying its own work with the full context and tools available to a human engineer, rather than being limited by missing context.

Workflow diagram · grounded in source
1
User submits prompt
trigger
“You can chat with Inspect in Slack and send it screenshots, use the Chrome extension to highlight specific changes to elements, prompt it on the web interface, discuss on the Pull Request”
2
Sandbox environment spins up
integration
“Each session runs in a sandboxed VM on Modal with everything an engineer would have locally: Vite, Postgres, Temporal, the works. It's wired into Sentry, Datadog, LaunchDarkly, Braintrust, GitHub, Slack, and Buildkite.”
3
Classifier routes to repository
routing
“build a classifier to determine what repository to work in. Take the user's incoming message, any thread context (if sent in a thread), and the channel's name”
4
Agent writes code
ai_action
“Inspect writes the code like any other coding agent”
5
Agent verifies its work
validation
“For backend work, it can run tests, review telemetry, and query feature flags. For frontend, it visually verifies its work and gives users screenshots and live previews.”
6
Pull request opened on GitHub
output
“The API will then use the user's GitHub token to call GitHub's pull request API”
Reported outcome

Inspect now writes approximately 30% of all pull requests merged to Ramp's frontend and backend repos, reaching that level of adoption within a couple of months without forcing anyone to use it.

Reported metrics
pull requests written by Inspect~30%
Time to reach adoption levela couple months
Reported stack
ModalVitePostgresTemporalSentryDatadogLaunchDarklyBraintrustGitHubSlackBuildkiteOpenCodeDurable ObjectsAgents SDKGPT 5.2
Source
https://builders.ramp.com/post/why-we-built-our-background-agent
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Inspect now writes approximately 30% of all pull requests merged to Ramp's frontend and backend repos, reaching that level of adoption within a couple of months without forcing anyone to use it.

What tools did this team use?

Modal, Vite, Postgres, Temporal, Sentry, Datadog, LaunchDarkly, Braintrust, GitHub, Slack.

What results were reported?

pull requests written by Inspect: ~30%; Time to reach adoption level: a couple months (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

User submits prompt → Sandbox environment spins up → Classifier routes to repository → Agent writes code → Agent verifies its work → Pull request opened on GitHub.

Related quality assurance cases