The agentic flywheel is, in Humans and Agents in Software Engineering Loops, the level beyond Humans on the loop: instead of improving the agent’s harness by hand, humans direct agents to manage and improve it. Agents are given the information they need to evaluate how the loop performs, review the results of each workflow step, and recommend improvements to any upstream part of the workflow that could improve those results — so that, in the article’s words, the agent harness “generates recommendations for improving itself.”
Usage
The article’s starting point for the signals is the tests and evaluations already in the harness. It suggests the flywheel becomes more powerful with richer signals: pipeline stages that measure performance and validate failure scenarios, and operational data from production, user journey logs and commercial results.
When It Applies
The article describes adopting it gradually. Humans first consider the recommendations interactively and prompt agents to implement specific changes, or have agents add recommendations to the product backlog to be prioritised, scheduled, and then applied and tested by agents in the automated flow. As confidence grows, agents can score their recommendations for risks, costs and benefits, and recommendations with certain scores might be approved and applied automatically. The author suspects that for standard, frequently done work this could end up looking much like humans out of the loop, once the improvement loops reach diminishing returns, but argues that it yields robust, “maybe even anti-fragile” systems that continuously improve themselves. This is a forward-looking proposal by one author, not a reported measured result.