Why We Built Our Own Background Agent

Why It Matters

Inspect demonstrates how tightly integrated AI agents can accelerate software delivery and reduce developer friction, setting a new benchmark for enterprise coding assistants.

Key Takeaways

• Inspect runs in Modal sandboxes with full dev stack
• Unlimited concurrent sessions cost near zero
• 30% of Ramp PRs now generated by Inspect
• Multiplayer sessions sync across Slack, web, VS Code

Pulse Analysis

Background coding agents have struggled with missing context, forcing developers to juggle local environments and AI prompts. Ramp’s Inspect tackles this by provisioning a complete development sandbox on demand, mirroring a developer’s machine with Vite, Postgres, Temporal and more. By embedding telemetry, feature‑flag checks, and visual previews directly into the agent, it closes the verification loop that typical LLM‑based tools leave open, delivering code that’s ready to merge.

The technical backbone relies on Modal’s instant‑start sandboxes and Cloudflare Durable Objects for stateful, real‑time collaboration. Each session lives in an isolated VM, pre‑built every thirty minutes to keep repositories fresh, while the API streams token updates via Cloudflare’s Agents SDK. This architecture enables true multiplayer editing—any teammate can join a session, see live changes, and retain authorship metadata, a capability rarely seen in competing products.

Business impact is immediate: Inspect now writes about a third of Ramp’s merged pull requests, slashing cycle time and freeing engineers from environment setup. The model‑agnostic design supports frontier LLMs and custom tools, allowing teams to experiment with new AI capabilities without re‑architecting infrastructure. As other firms watch Ramp’s rapid adoption, the bar is rising for AI‑driven development platforms that combine speed, context richness, and collaborative flexibility.