Elastic Loop defines tight and loose AI work cycles to match task criticality.
Tight or loose is a design choice per task. Same person, same day, two different modes.
What It Is
How intensely do you work with the AI? Robert Glaser (code.talks 2025) sorts this into two modes at the same tool, not as a character trait, but as a deliberate setting per task.
Tight loop:
Short cycles: prompt → output → immediate check → follow-up → output. You stay in the loop, every intermediate step is verified. Appropriate for production code, legally binding text, safety-critical tasks, anything near the Jagged Frontier. High quality, heavy cognitive load on the human.
Loose loop:
Long cycles: brief → AI runs autonomously → interim results come in → you sort. You step back, give the AI room. Appropriate for research, exploring options, early concept phases, tasks deep inside the frontier. Less micromanagement, higher risk of drifting in the wrong direction.
The loop is elastic, not rigid - tightened or loosened per task. Which length fits depends on where the task sits relative to the frontier: near the edge, tight is required. Deep inside, loose is safe. Outside the frontier, don't delegate at all.
How To Spot It
- Someone shifts deliberately between tight and loose during the day, depending on the task
- Critical work is forced into short cycles, exploratory work is allowed to run long
- The person notices which mode they're in, no default tunnel
- The team talks about loop length, not just about output
What To Do (FL1 - Team Level)
- Set it deliberately: Decide before the task which loop length fits. The default mode (usually tight, out of mistrust) is rarely the right one
- Couple to the frontier: Loop length follows the frontier, not your mood
- Make it visible: Talking about loop length helps choose it deliberately. "I'll let the AI run for an hour" is a legitimate statement, not a laziness indicator
- Combine with Cognitive Load Shift: Tight = lots of review load, loose = lots of orchestration load. Both cost, just differently
The Trap
Sticking to the default mode. Always tight-looping wastes the AI's exploratory strength. Always loose-looping risks output nobody really inspected. The elasticity is the point, lose it, and the tool is half used.