Vercel open-sources durable-harness for crash-proof coding agents

Vercel open-sourced durable-harness on 21 October, a set of adapters that run coding-agent harnesses on top of the Workflow SDK’s durable backend rather than as a plain long-lived process. Supported harnesses are Claude Code, Codex, OpenCode, Pi and fx, each wired in through its own subpath import.

Every session becomes a workflow run: the conversation, the turn in flight, and every model and tool call land in that run’s event log. If the process crashes or a function times out, a new function reads the log and resumes the turn, including on a different machine. Completed model calls are never re-executed, so a crash doesn’t duplicate token billing. Tools have to opt in to being re-run safely:

tool({ replay: 'safe' })

Anything not marked that way comes back to the model flagged as “possibly run” instead of being retried blindly. Deployment options include Vercel itself, a file-based mode for local development, or self-hosting against Postgres, MongoDB, Redis or Turso; Claude Code and Codex sessions also need a Vercel Sandbox for their files. The project is explicitly experimental, versioned 0.0.x, with no API stability guarantee yet.

Why it matters: anyone running coding agents for tasks that take hours has had to build their own resume-after-crash logic, or just accept losing the session. This gives that a shared, open-source shape across five harnesses at once, which matters more than any single feature once agents are expected to run unattended.

The caveat: it’s a 0.0.x release, so treat the API as unstable until it settles.