Loop engineering
Designing the system that drives an agent over time, instead of typing every prompt by hand. A loop is a small program that re-runs the agent on a trigger (schedule, event, condition), with memory, an evaluation of the result, and orchestration. What separates it from a classic scheduled job, the cron that has been around for decades, happens when it wakes up: a cron always runs the same command, a loop is handed an outcome to reach and decides for itself what to do to get there, within the limits it was given. The leverage point shifts: you no longer write the right prompt, you design the loop that decides what to ask, when, and whether the output is acceptable. Popularized mid-2026 by Boris Cherny (Head of Claude Code) and Peter Steinberger (creator of OpenClaw, who joined OpenAI in February 2026), echoed by Addy Osmani.
Strengths
- Moves the effort from one-shot to the system: trigger, memory, eval, orchestration
- Makes agents useful continuously (scheduled) instead of reactive to each request
- A clear frame to reason about cost, guardrails and stop conditions
Limitations
- The term rides a 'prompting is dead' slogan more catchy than true: the prompt still lives inside the loop
- A poorly bounded loop (no eval, no guardrail) repeats and amplifies its mistakes
Best for
- A tech lead orchestrating agents in production who wants a design frame
- A PM who wants to grasp where the work goes when AI moves from assistant to process