Ramp Inspect vs Stripe Minions
What is being compared
Two internal unattended coding-agent systems:
Comparison table
| Dimension | Ramp Inspect | Stripe Minions |
|---|---|---|
| Primary infrastructure story | Hosted on modal sandboxes and distributed primitives | Runs on Stripe devboxes and internal infrastructure |
| Main positioning | Full-context background coding agent accessible to many kinds of builders | One-shot end-to-end coding agent optimized for Stripe engineers at massive scale |
| Entry points | Slack, screenshots, Chrome extension, web UI, PR discussion, web VS Code | Slack, web/CLI, internal docs, feature flag UI, ticketing integrations |
| Verification emphasis | Browser verification, telemetry, feature flags, screenshots, previews | Deterministic local checks, curated tools, bounded CI, strong internal rule files |
| Orchestration idea | Agent in a rich sandbox with plugins, snapshots, and child sessions | Blueprint abstraction mixing deterministic nodes with agentic nodes |
| Collaboration model | Strong emphasis on multiplayer collaborative sessions | Strong emphasis on unattended runs handed back for review |
| Core adoption claim | ~30% to over half of merged PRs depending on article / timeframe | 1,000+ then 1,300+ merged PRs per week |
Main synthesis
Ramp and Stripe are solving similar problems but from different starting points.
Ramp’s articles emphasize product experience: speed, rich interfaces, broad accessibility, browser-based verification, and multiplayer workflows. The system is presented almost as a shared coding surface for the whole company.
Stripe’s articles emphasize systems design and reliability: large-codebase constraints, deterministic orchestration, scoped context, internal tool reuse, and bounded testing loops. The system is presented as a carefully engineered extension of existing developer productivity infrastructure.
Takeaway
Both examples support the same high-level thesis in background-coding-agents: unattended agents work best when they inherit the real tools, environments, and constraints of the engineering organization rather than operating as generic standalone copilots.