Knowledge Management for Coding Agents

Abstract

An agent makes an architectural decision, the session ends, and the repository holds the result without any underlying reasoning. Three months later, someone looks for the rationale and finds the commit. This talk highlights the problems this creates in ongoing projects, ranging from onboarding and code reviews to the question of why an agent proposes the exact same rejected solution for the third time. After that, we explore the countermeasures: documented decisions, logged tool calls, and a wiki that agents can read and write themselves.