New Odyssey

Outcomes · For engineering teams

Agents that finish the job after a failure

When a worker dies mid-run, the agent resumes where it stopped and reuses the steps it already completed.

Who it is for

Who this is for

Engineering teams running multi-step agent workflows that wait on approvals, background jobs or flaky providers.

What changes

Before and after

Before

  • A crashed run starts again from the beginning, or is abandoned.
  • Restarting repeats work that already happened, including customer updates.
  • A timeout is treated as a failure, and the write is sent twice.

After

  • Each step's result is saved, along with the step that should run next.
  • A replacement worker restarts with the same run ID, reuses the completed work and continues.
  • Progress survives approval waits, background jobs and crashes.

How it is delivered

Delivered by a product, not a project

Delivered by Romulus, through an API and a JavaScript SDK.

  • Execute the next step, checkpoint its result, wait or be interrupted, then resume.
  • The API runs locally and keeps its state in SQLite.

What you provide

What we need from you

  • Node.js 24 to run the preview.
  • A workflow you can express as steps with a result worth saving.
  • Idempotency or reconciliation for your external writes.

How the result is measured

How you will know it worked

  • Run the quickstart: it kills a worker mid-step, restarts it and lets you inspect the recovered run in SQLite.
  • For your own workflow, check that completed steps are not repeated after a forced crash.

How to start

How to start

Run the Romulus quickstartRead the docs

Romulus runs on your machine today. There is no hosted service and no account to create.

Good to know

Limits, stated plainly

  • Completed steps are reused. Unfinished work may run again, so external writes still need idempotency or reconciliation.
  • A timeout does not mean a write failed: check the original operation with the provider before retrying.
  • This is a development preview that runs locally, not a hosted production service.

For engineering teams

Other outcomes