Vercel Open-Sources durable-harness for Crash-Proof Coding Agents
Oct 21, 2026, 3:15 PM · MarkTechPost
Vercel is going after the boring failure that kills cloud coding agents: a function dies mid-turn and the work, and the money spent on it, evaporates.
Why it matters
Coding agents such as Claude Code, Codex, OpenCode and Pi were built as long-running processes on a developer's laptop, holding the session in memory. Move them into short-lived cloud functions and a timeout or network error can wipe out a turn. Vercel's new open-source package, durable-harness, wraps those existing harnesses so each step lands in a durable log and a crashed turn picks up on a fresh function.
This matters because the industry is shifting agents from personal machines to background cloud workers. Reliability, not model quality, is often what decides whether that shift sticks.
From the desk
We like the shape of this. Rather than asking teams to rewrite their agents in a new framework, Vercel built adapters that leave the harness alone and add durability underneath. The core idea is old and sound: record an intent before a model or tool call, record the outcome after, and on recovery read finished work back instead of redoing it. That means completed model calls are not billed twice, which is real money for anyone running agents at volume.
The smartest detail is the replay setting on tools. Lookups can be marked safe to rerun. Actions like refunds can be marked never, and the model is told the call may already have happened. That is exactly the kind of guardrail agents need as they start touching payments and production systems, and I'd like to see it become standard vocabulary across agent frameworks.
Now the caveats, and they are significant. This is version 0.0.x and labeled experimental. There are no published latency or throughput benchmarks, and the recovery claims come from Vercel itself. Claude Code and Codex only run inside a Vercel Sandbox, even when self-hosting. So the portability story is partial: Pi and OpenCode travel freely, the two most popular commercial harnesses less so.
There is also a platform play here. The package rides on Vercel's Workflow SDK and works best on Vercel Functions and Queues, with every session traced in its Agent Runs view. A self-hosted Postgres path exists and takes two environment variables, which is a fair concession. But the likelier outcome is that durable-harness becomes an on-ramp to Vercel's agent hosting, and teams should evaluate it with that in mind.
If durability like this becomes table stakes, the bigger effect is cultural. Agents that survive crashes get trusted with longer jobs. That is useful, and it raises the bar on audit logs and approval gates, because a resilient agent doing the wrong thing keeps doing it.
Context
durable-harness ships as an Apache-2.0 npm package with adapters for Pi, OpenCode, fx, Claude Code and Codex. Workspaces can be a virtual file system stored in the session log (about 500 files, no network), in-memory scratch space, or a per-session Vercel Sandbox. It can run on Vercel, on local files, or self-hosted against a Postgres backend; Vercel also lists other storage options. Comparable durability approaches include Temporal with the OpenAI Agents SDK, Inngest AgentKit and Cloudflare's Agents SDK.
Who feels it
- Developers
- A drop-in way to stop losing agent turns and paying twice for finished model calls, with explicit replay rules for risky tools. Expect API churn at 0.0.x.
- Platform teams
- The self-hosted Postgres option lowers lock-in, but Claude Code and Codex still require a Vercel Sandbox, which shapes architecture choices.
- Competing infra vendors
- Temporal, Inngest and Cloudflare now face a vendor that wraps popular harnesses directly rather than asking teams to adopt a new framework.
What to watch
- Independent benchmarks on recovery overhead and latency
- Whether Claude Code and Codex support expands beyond Vercel Sandbox
- A move from 0.0.x toward a stable API
- Adoption of replay-safe and replay-never conventions in other agent frameworks
Companies: Anthropic