Vercel has released durable-harness, an open-source package designed to keep coding agent sessions alive through crashes and timeouts. The new tool wraps existing harnesses like Pi, OpenCode, Claude Code, and Codex without changing them, storing each session in a durable log so that a dead function no longer means lost work for developers moving agents into the cloud.
The package records an intent before each model or tool call and logs the outcome once the function finishes. Backed by Vercel Queues, the system ensures that if a function dies during a turn, a new function can read the log and continue seamlessly. Completed model and tool calls are read back from the log rather than re-run, meaning finished tokens are never billed twice.
The software supports multiple workspace options, allowing users to run agents on Vercel, locally, or self-hosted using Node.js and a Postgres World. While the experimental 0.0.x API currently lacks published benchmarks and requires a Vercel Sandbox for tools like Claude Code and Codex, it offers a robust solution for cloud-based development.
Why it matters: This open-source release closes a real gap for teams moving AI coding agents off local laptops and into resilient cloud environments.