Self-Learning in Enterprise Agents

What the series argues

A self-learning agent in a financial system is safe in proportion to how little it learns. The appeal of an agent that improves itself is obvious. The risk is that it changes behaviour nobody signed off, inside a system where every number is defensible or it is worthless.

The first paper sets the discipline of what not to learn, and why a production agent absorbs a narrow slice, governs it by hand, and never rewrites its own reasoning. The second covers configuration, how a learned rule hangs at a level in the business's own hierarchies and inherits down, where it is overridden, and the close cadence that decides what enters the rule store. The third makes the uncomfortable argument. Most of what an agent learns is a deferred fix for process, data or technical debt, which makes the rule store a map of everything you have not repaired.

Read the third paper before you commission a learning agent.