Vercel's durable-harness Enables Automatic Crash Recovery for AI Agents

Introduction

Vercel has released durable-harness, an open-source project that aims to eliminate the scenario where an AI coding agent loses all progress when the host machine crashes or loses power. The project adapts Vercel's internal durable-execution infrastructure — originally built for the eve agent and the Workflow SDK — so that existing coding-agent runtimes (called "harnesses") can persist every step to a durable log and resume seamlessly after a failure.

Supported Harnesses

Directly compatible: Pi, fx, eve — these are designed for embedding and work out of the box.

Require an extra sandbox: Claude Code, Codex, OpenCode — because they were not built for embedding, they need an additional sandbox layer to participate in the durable log.

Core Architecture: Sessions as Workflow Runs

Each agent session becomes a single Workflow SDK run. Every model invocation and every tool call is recorded in two phases:

Intent — the planned action is written to the log.

Outcome — the result (or error) is appended to the same log entry.

All entries go into one durable log. If the running function dies, a queue hands the turn to a fresh function instance, which reads the log and continues from the last completed step. Because completed call results are fetched directly from the log, the same model call is never paid for twice.

Tool Replay Strategies

The replay policy for tool calls is a critical design point:

Safe tools (e.g., reading files, checking status) are automatically re-executed after a crash.

Never tools (e.g., issuing refunds, sending notifications) are never run a second time. The model is simply informed that "this call may have already happened."

This distinction prevents duplicate side-effects while allowing idempotent operations to retry transparently.

Deployment Flexibility

The persistence backend is configurable via two environment variables:

Local development: default file-based storage.

Vercel-hosted: uses Vercel's managed infrastructure.

Self-hosted: point to a Postgres database.

Switching between these targets requires no code changes — only the environment variables — so the entire agent can move from Vercel's cloud to a private Postgres instance while keeping the same durable log semantics.

Filesystem Modes

Three filesystem abstractions are provided to match different workload needs: virtual() — files are stored inside the durable log; suitable for short-lived tasks. virtual({durable:false}) — in-memory temporary files that are discarded at the end of each turn. sandbox() — provisions an independent Vercel Sandbox with /workspace backed by Vercel Drive; intended for real-world projects that need a persistent, isolated workspace.

Documents: Structured Agent Memory

Beyond chat history, durable-harness introduces Documents — typed JSON structures (to-do lists, plans, etc.) that are stored alongside the log and shared with the application. This expands the agent's "memory" to include workspace files and task-tracking artifacts, not just conversation context.

Open Source & Community Feedback

The project is built on the Workflow SDK 's World interface and is available on GitHub: https://github.com/vercel-labs/durable-harness.

A community member noted that recoverable harnesses have existed before, but durable-harness distinguishes itself by making the entire compute layer ephemeral — all state lives in the log — enabling true cross-machine recovery.