WTF is Agentic Flywheel?
Agents improving their own harness, using feedback from tests, pipelines and production to recommend changes.
TL;DR
Harness engineering means humans fix the environment whenever an agent makes a mistake. The flywheel goes one step further: agents analyse signals such as failing tests, CI results and production data, and propose the harness improvements themselves, with humans directing and approving. Each turn makes the next turn better.
How it works
Described by Kief Morris as humans directing agents to manage and improve the harness itself, using feedback signals from tests, pipelines and production to generate continuous improvement recommendations. It keeps humans on the loop rather than in it, and depends on trustworthy signals; a flywheel fed by flaky tests amplifies noise.
Origin
- Coined / introduced by
- Kief Morris (Thoughtworks)
- When
- 4 Mar 2026
- Story
- Introduced in 'Humans and Agents in Software Engineering Loops'.