Vercel's durable-harness makes Claude Code, Codex and Pi survive crashes

Why harnesses need durability in the cloud

Developers commonly run coding agents locally through harnesses like Pi and Claude Code. Vercel says keeping such agents working around the clock requires a computer that stays on and awake, which led to the laptop lid memes and a spike in Mac mini demand. Harness workloads are now moving to the cloud, according to the company.

Vercel points to individual developers deploying always-on coding agent workflows, to projects like OpenWork and HanaAgent that use open-source harnesses as the engines behind their apps, and to engineers building their own harness products. Agent tasks often run for minutes or hours.

Locally, a harness has continuous access to the workspace and session history on the same disk, so resuming interrupted work takes a simple prompt. In the cloud, a VM with persistent storage can replicate that setup, but Vercel notes that the VM does not record which model and tool calls completed or resume interrupted turns after a crash. A failed turn or network error means lost work.

How durable-harness works

Vercel already built eve and fx to run durably in the cloud. durable-harness extends this to other harnesses and uses the same machinery that powers Workflow SDK. It is a set of adapters and utilities that store a harness session in a World, the Workflow SDK's durable backend.

The package writes each step of a turn to a durable log. Before a model call or tool call runs, it records an intent, and when the call finishes, it records the outcome. The turn is also backstopped by a queue message. If the function dies, the queue hands the turn to a new function, which reads the log and continues.

• Completed calls never run again, because their results come from the log. Vercel says users are never billed twice for the tokens of a finished model call.
• Interrupted tools marked replay: 'safe' run again.
• Any other tool is reported to the model as possibly run, and the model decides whether to check.

Harnesses that keep their own session files, such as Claude Code's transcript, Codex's rollout and OpenCode's event store, have every record mirrored into the log. A new function or Sandbox can rebuild them from it. For fx and eve, durability is built in, while for the rest durable-harness adds it from the outside without changing the harness.

Tools, files and where it runs

Each tool declares what should happen if a crash interrupts it. A 'safe' tool, such as a lookup, can run again. A 'never' tool, such as a refund, never runs twice, and the model is told the call may have happened so it can check first.

For harnesses that need a shell and files, Vercel says running them directly on a shared serverless function host is not safe. durable-harness offers three options:

• virtual() keeps files in the session's log, with a bash-compatible shell that cannot reach the function's files or the network. It suits short tasks and small workspaces.
• virtual({ durable: false }) keeps files in memory for scratch work and clears them after every turn.
• sandbox() gives the session its own Vercel Sandbox, with /workspace on a Drive that outlives every sandbox and function. It suits real projects and long-running work.

Because the package writes to the World interface, the same agent can move between backends without code changes. On Vercel it uses the Vercel World, in development it uses files on the local machine, and on self-hosted servers it can use Postgres, MongoDB, Redis or Turso. Two environment variables, WORKFLOW_TARGET_WORLD and WORKFLOW_POSTGRES_URL, move every agent using durable-harness to Postgres.

What it means for developers

Developers can run supported harnesses durably on Vercel, self-host with their own Postgres instance, make a local harness workflow durable, or build their own durable harness. Vercel's Agent Runs view shows the trace of every session, including each model and tool call, and turns that resumed after a crash appear in the same trace so users can see where one stopped and picked back up.

For teams running long agent tasks in the cloud, the practical effect described by Vercel is that a timeout or dropped connection no longer forces a manual restart or repeated work. Setup guides for each harness are in the Vercel docs.