← Blog
·4 min read

Loop engineering: the skill isn’t prompting anymore

The shift from writing clever prompts to designing self-correcting loops — and why every good loop needs a queue to draw from and a stop condition to end on.

For two years the implicit skill of working with AI was prompting: find the right words, and the model does the thing. In 2026 that quietly stopped being the frontier. The new skill, as a wave of practitioners started naming it, is loop engineering — designing the self-correcting loops that run the agent, not the single prompt you hand it.

Addy Osmani framed the shift directly: a loop is a recursive goal — “triage every open issue and draft fixes” — that keeps acting and checking its own results against real signals (tests, type-checkers, linters) rather than against your patience. You stop steering each step and start designing the system that steers itself.

What actually makes a loop work

The reason this is an engineering skill and not a prompting trick is that good loops have structure. Two parts matter most, and both are easy to forget in the excitement of “it just keeps going”:

  • A queue to draw from. A loop needs a source of work. “Keep improving things” has no queue, so it has no shape; “work through these tasks” does. The queue is what turns a loop from a vibe into a process.
  • A stop condition. A loop without an off switch is the billing-disaster story everyone’s heard. The cleanest stop condition is an empty queue: when there’s nothing assigned, the loop idles.

Get those two right and the loop becomes boring in the best way — it picks up work, does it, verifies, and stops when the work runs out. Get them wrong and you get either a loop that never starts anything useful or one that never stops.

The loop needs somewhere to look

Here’s the part that connects loop engineering to a humble to-do app: every durable loop needs a place to read its work from and write its results to. People bolt this together with queue tables, issue trackers, and markdown files — but the thing they’re reinventing is a task list with an API.

That’s the role Lume is built to play. Its MCP server and REST API give a loop a clean place to ask “what’s assigned to me?”, do each task, and post the result back as “ready for review.” The queue is your real task list; the stop condition is an empty one; the review gate is the human checkpoint. Read the Mac mini loop walkthrough for a concrete build.

Prompting got us to capable agents. Loop engineering is how you put them to work without babysitting them — and a bounded, reviewable queue is the part of the loop most people are still missing.

Want a list your agents can pull from?

Lume gives every task an API, an MCP server, and an assignee. Free to start.