Why AI Coding Agents Need Work Attempts, Leases, and Checkpoints

The article argues that conventional task assignment breaks down when AI coding agents crash, forget, or overlap on the same work.

It highlights rhizome-mcp’s model, where an agent claims an issue as a leased work attempt that must be renewed to stay active.

When the lease expires, the task is automatically released, reducing stale in-progress states and duplicate effort.

Checkpoints preserve progress safely, making the workflow more resilient for unreliable agents.

The main point: AI agent coordination needs distributed-systems-style primitives, not just standard issue trackers like Jira, Linear, or GitHub Issues.