Vercel has released durable-harness, an open-source package that lets a coding agent survive a crash in the middle of a job and carry on without starting over. It works with four existing tools: Pi, OpenCode, Codex, and Claude Code. Vercel’s engineering blog, in a post by Pranay Prakash, describes how it works.
Start with the vocabulary. A harness is the software wrapped around a coding AI model. It runs the loop in which the model reads files, calls tools, and decides what to do next, and it holds the session together. Developers have mostly run harnesses on their own laptops, which means the laptop has to stay open and awake for the agent to keep working. Vercel’s post jokes about the “laptop lid memes” and the run on Mac minis that followed.
Moving a harness to the cloud solves the lid problem and creates a new one. On a laptop, the session history and the project sit on the same disk, so an interrupted job can be resumed with a short prompt. A cloud virtual machine with storage can copy that setup, but it does not record which model calls and tool calls had already finished. When a function times out or the network drops, the work in flight is lost.
Durability is the fix, and Vercel means something specific by it. The package keeps a running log of every step in a turn. Ahead of each model or tool call it records the intent, and when the call returns it records the result. If the process dies, a queue message hands the turn to a fresh function, which reads the log and continues from the first unfinished step.
The practical gain is money and time. Vercel says finished model calls are never run a second time, so you are not billed twice for the same tokens. Tools that were cut off mid-run follow a rule the developer sets. A tool marked safe, such as a file lookup or a check on CI status, simply runs again. A tool marked never, such as a refund, does not. Instead, the model hears that the call might already have gone through, and it decides whether to check first.
Claude Code, Codex, and OpenCode keep their own session files, so durable-harness mirrors every record of those into the log and rebuilds them on whatever machine picks up the job. Vercel’s own agents, fx and eve, were already built to be durable.
The self-host option matters for anyone wary of tying agent infrastructure to one vendor. The package writes to Workflow SDK’s “World” interface, which bundles the step log, the crash recovery queue, and the live feed of what the agent is doing. On Vercel it uses Vercel’s own World. In development it uses local files. On your own servers it can use Postgres, MongoDB, Redis, or Turso. Moving a Postgres setup takes two environment variables, according to the post.
Vercel’s post is a product announcement, and it offers no failure-rate or recovery-time figures. The demo it describes compares an agent with and without durability, but that is the company’s own test.
Long agent jobs that touch real systems are where this will show up first. Teams running coding agents against production repositories or payment tools should check how their current harness handles a mid-turn crash, because a retry that repeats a refund or a pull request comment costs more than the lost minutes.
Reported by Vercel’s blog in a post by Pranay Prakash, published in October 2026.
