it worked, but was hard to tell what agents were doing, why something failed, or whether a workflow was actually progressing.

We thought it would be more interesting to treat agents as long-lived workers with state and responsibilities and explicit handoffs. Something you can actually see and reason about, instead of just tailing logs.

So we built Clawe, a small coordination layer on top of OpenClaw that lets agent workflows run, pause, retry, and hand control back to a human at specific points.

This started as an experiment in how agent systems might feel to operate, but we're starting to see real potential for it, especially for content review and maintenance workflows in marketing. Curious what abstractions make sense, what feels unnecessary, and what breaks first.

Repo: https://github.com/getclawe/clawe