The Mikado Method (attempt a change naively, map what blocks you as a graph, undo, work the leaves first) has always been a sound way to approach large refactors. The reason it doesn't stick is the undo step: discarding hours of work is hard for humans under deadline pressure. AI coding agents invert that cost. A discarded branch costs seconds. This talk covers how to apply the method when the agent handles the attempts and you maintain the graph.

The human's job changes when agents do the implementation. Prerequisite discovery and graph maintenance become the primary skill: writing specs that produce useful failures, turning those failures into graph nodes, and assigning leaf tasks to fresh agent sessions.

This talk walks through how to apply Mikado thinking to AI-assisted development: how to write specs that produce useful failures, how to turn those failures into graph nodes, how to assign leaf-node tasks to fresh agent sessions, and how to know when a branch is done. It also covers the limits: where the graph becomes too expensive to maintain and when a simpler approach is the right call.

Audience: developers and technical leads using AI coding agents (Copilot, Cursor, Claude Code) for non-trivial implementation work. No prior knowledge of the Mikado Method required.

Duration: 25–30 min. adaptable to a lightning talk

First delivery. No special requirements.