Vercel Open Sources durable harness for Crash Proof Coding Agents
Jessica and Cathy dig into Vercel's durable-harness, an open-source layer that makes coding agent sessions survive crashes by logging every turn, intent, and outcome into a durable workflow world. They focus on the real product gap it fills for cloud-hosted agents, the adapter strategy that wraps existing harnesses without rewriting them, and the trade-offs around state, storage, and portability.
Transcript
Jessica The annoying part isn’t the model. It’s that your coding agent can be halfway through a turn, the function dies, and the work just evaporates. Vercel’s durable-harness is basically saying, no, keep the session alive.
Cathy Right, and that’s the actual cloud problem. Local harnesses get away with living in memory because the disk and the session are sitting in the same place, but once you move that into a function, you need a record of what already happened.
Jessica Exactly. It wraps Pi, OpenCode, Claude Code, and Codex without asking them to become a different product, which is the part I like. You don’t have to rebuild the agent to get crash recovery.
Cathy Mm-hm.
Jessica It also fits the Vercel shape pretty cleanly. Each session lives in a durable log, so when a turn is interrupted, the next function can read what finished, what was in flight, and what never started.
Cathy That’s the useful bit. They’re not pretending the function is stateful, they’re making state external and replayable, which is the only sane move if you want recovery after a timeout or network error.
Jessica And the mechanics are pretty concrete. Before a model call or a tool call runs, it records intent. When it comes back, it records the outcome. So you get a turn history that isn’t just a transcript, it’s execution history.
Cathy Exactly. That’s also why the queue matters. Each turn is backed by a Vercel Queue message, so a dead function hands off the job instead of taking the session down with it.
Jessica I did the reading, but not a full fact-check rabbit hole on this one, so if one of these details sounds weirdly exciting, go look it up. This field loves moving the furniture around while you’re still sitting on the couch.
Cathy Fair. And the architecture is basically World-plus-log: storage, queue, and stream behind one interface, then the harness sits on top of that. It’s neat because the harness doesn’t need to know whether it’s running on a laptop or in a managed world.
Jessica That’s the product story, honestly. If you’re already shipping an agent workflow, the missing piece is not another clever prompt loop. It’s making sure the loop survives the boring failures that happen in production.
Cathy Sure, but I’d keep one eyebrow up. Moving the state into a durable runtime is elegant, but it also means your harness now lives and dies by the storage and replay semantics. That’s a real trade-off, not free magic.
Jessica No, totally. But compared with losing a half-completed coding task because a function blipped, I think most teams will take the trade. Especially if they’re already on Vercel and don’t want to rewrite the whole harness stack.
Cathy The mirrored session files are what make that believable. Claude Code’s transcript, Codex’s rollout, OpenCode’s event store — those all land in the log too, so you’re not just preserving one narrow interface.
Jessica And there’s typed JSON state for little structured bits, like a to-do list, capped at sixty-four kilobytes. That’s a nice sign they’re thinking about the messy middle, not just the happy-path transcript.
Cathy Mm. It’s also a good reminder that this is infrastructure for a specific workflow. It’s not trying to be the whole agent product, just the part that stops work from disappearing when the runtime gets bored and dies.
Jessica Which is why I think the target user is obvious. It’s teams moving coding agents off laptops and into cloud runs, where resumability is no longer cute, it’s the difference between usable and annoying.
Cathy Yeah, and the open-source angle matters because it lowers the migration friction. If you already have a harness, wrapping it beats a rewrite. That’s the path that actually ships.
Jessica And if you want to try it, the repo is live and they point to the docs and the announcement. The nice thing is you don’t have to invent a new mental model. You just give the session somewhere durable to land.
Cathy That’s the part I buy. Not because it’s glamorous, but because it’s the missing plumbing for a class of agents that were never really built for the cloud in the first place.
Jessica Come on, Cathy, that’s almost a compliment.
Cathy Don’t get used to it.
Jessica Fair. Anyway, this one feels like the kind of thing we’ll keep seeing under different names, and we’ll pretend we were surprised later.