Context-Rich Sandboxes Beat Generic Coding Agents

·https://builders.ramp.com/post/why-we-built-our-background-agent
articleaiagentscoding-agentsopencodeinfrastructuresandboxesagent-loops

Maps directly to opencode-vibe: hosted OpenCode sessions with warmed sandboxes, synced clients, verification tooling, and multiplayer handoff.

Joel dropped Ramp’s background agent post in Slack as “the shape of thing i have in my head for opencode-vibe.” The useful bit isn’t that Ramp built another coding agent. It’s that Ramp Inspect is company-specific developer infrastructure: OpenCode running inside warmed Modal sandboxes, wired into Sentry, Datadog, LaunchDarkly, Braintrust, GitHub, Slack, and Buildkite with enough context to verify work like a Ramp engineer.

This is clever because it moves the bottleneck from setup to model intelligence. Ramp prebuilds repo images every 30 minutes, restores file-system snapshots with Modal Sandboxes, lets the agent read before sync completes, blocks writes until sync is done, queues follow-up prompts, and keeps a stop path. That’s the background-agent shape: fast start, isolated execution, resumable state, and no laptop in the loop.

The part that maps hardest to joelclaw is the surface area. Slack, web, Chrome extension, GitHub pull request comments, hosted VS Code, screenshots, live previews, and multiplayer sessions are all clients on the same session state. The agent is not a terminal toy. It’s a shared work object that can show up wherever the work already happens.

The adoption claim is worth pressure-testing: Ramp says roughly 30% of merged pull requests in its frontend and backend repos were written by Inspect after a couple months, without forcing usage. Their bet is not “better chat.” It’s make the agent cheap enough, visible enough, and context-rich enough that people stop rationing it.

Key Ideas

  • Ramp Inspect treats background coding as internal developer infrastructure, not a thin wrapper around chat.
  • Modal Sandboxes and sandbox snapshots make sessions fast by prebuilding repo images, restoring state, and syncing only recent changes.
  • OpenCode matters because it is server-first, with a typed SDK and plugin system that make custom clients possible.
  • Cloudflare Durable Objects give each session its own SQLite state, while the Cloudflare Agents SDK handles real-time streaming across clients.
  • Multiplayer sessions keep authorship attached to prompts so product managers, designers, and engineers can all collaborate on the same coding session without collapsing review boundaries.
  • Verification is part of the loop: tests, telemetry, feature flags, screenshots, live previews, and pull request updates are tools the agent can use to prove work.
  • Public workflow surfaces like Slack create adoption loops because people see the agent working in shared channels instead of hidden local terminals.

Links

← All discoveries