Cloudflare Computer v0.7.0: replay after crashes, pause for approval mid-run

Cloudflare Computer v0.7.0 lets agent-run JavaScript resume after deploys or crashes and pause for human approval. For builders on Durable Objects.

Picture a two-person dev shop that built an agent to triage GitHub issues for a client. The agent writes small JavaScript programs that read the repo, draft a reply and post a comment. It works on a quiet afternoon. Then someone deploys mid-run, the run dies halfway through, and nobody knows which comments went out. The client also wants a human to sign off before anything gets posted publicly, so the team has bolted on a queue, a status table and a polling loop. Now they maintain an approval system as well as an agent.

That is the gap Cloudflare's changelog for @cloudflare/computer v0.7.0, dated October 26, 2026, targets. Code an agent runs in a Computer workspace can now "resume after a deploy or crash, and pause on an external action, for example to approve a tool call." The catch is in the same post: replay can repeat a call whose result wasn't recorded before the interruption.

What it is

Computer is Cloudflare's agent workspace library. According to the GitHub repo, it is "a virtual filesystem that lives inside a Durable Object," with state held in SQLite and code run through one execution surface, workspace.runtime. Three backends ship: a container, an isolate shell, and an isolate JavaScript backend that runs an ECMAScript module in a fresh Dynamic Worker. An outside overview from Flowtools describes it as infrastructure for your agent "rather than a ready-made personal assistant."

v0.7.0 adds two things, per the changelog:

Alternatives, each from its own source:

Who it's for

Pick this up if you already run agents on Workers and Durable Objects, use Computer's JavaScript backend, and let agent-written code call your functions that touch the outside world: posting comments, writing records, sending mail. You get crash recovery and an approval gate without adding Postgres or a workflow engine.

Skip it if you're not on Cloudflare. Skip it if your approval is one step in a longer workflow you already run on Temporal or Inngest. Skip it if you're on LangGraph with a checkpointer that works. And skip it if your agent mainly runs shell commands in the container backend, which the repo marks as experimental.

Before and after

Before, for the issue-triage agent:

  1. The agent writes a JavaScript module and it runs in the workspace.
  2. A deploy or crash kills the run. You dig through logs to find which calls finished.
  3. For approval, postComment writes a row to your own table and returns early. The run ends.
  4. A human approves in your UI. You re-trigger the agent and rebuild its context.
  5. You hope the re-run doesn't post twice.

After, as the changelog describes it:

  1. Set onInterrupt: "replay" on the backend.
  2. postComment in your custom module calls ctx.interrupt(). The run pauses there.
  3. Your UI lists workspace.runtime.pending() and calls respond() with the decision.
  4. The run continues once all pending calls are resolved, even if you deployed in between.
  5. If the run is interrupted, it replays from the top on the next workspace.runtime call, or from a Durable Object alarm calling recover(). Recorded results are reused.
  6. For calls that must never run twice, you still have to handle retries. The changelog points to its "Calls that must not run twice" page and unsafeToRepeat().

Try it

The Cloudflare Agents file system docs give the setup:

npm i @cloudflare/computer zod

Then set onInterrupt: "replay" and copy the changelog's postComment approval example as your first gated function. The repo's examples directory includes examples/worker-javascript, which mirrors the shell example but evaluates an ECMAScript module. The sources don't say which Workers plan this requires.

Where it breaks

Sources