Koestler AI

Koestler AI argues back. Write down the decision you are leaning toward, and it lays out your reasoning and attacks the weakest link, one question at a time. You answer, concede, or leave the question standing on the record. At the end you get a delta: the position you walked in with, the one that survived, and why it changed.

The same core checks what AI agents do. Connected over MCP (Claude Code today), it judges consequential actions such as file writes, shell commands and deploys against an explicit rule table before they run. Each is passed, debated, handed to you, or blocked; nothing is silently approved. By design, the judge comes from a different model family than the agent, and every verdict goes into a record that cannot be edited. It starts in shadow mode, which blocks nothing and records everything.

The name comes from Arthur Koestler's idea of bisociation: creativity happens where two unrelated frames of thought cross. Its insight engine borrows ideas from Copycat, the model of analogy-making built by Douglas Hofstadter and Melanie Mitchell. When it crosses a stuck problem with a distant field, the result stays labeled a hypothesis until tested.

Koestler AI is in beta. We do not yet claim the gate reduces accidents. That claim waits on a controlled trial with real agents.