Definition
Plain language
A method for spotting which member of an AI agent team caused a failure, by tracing blame backward through who influenced whom, then rewriting that member's instructions.
As stated in the literature
Gradient-Based Connections — weights each inter-agent edge by the gradient sensitivity of a downstream agent's output to an upstream agent's text, prunes to a blame graph, and traces a verbal loss backward to localize and repair the responsible agent's prompt.
Also called: Gradient-Based Connections
Why it matters: It matters because pinpointing which agent caused a multi-step failure lets you fix the real culprit instead of blaming the last one to speak.
For example, when a team of agents botches a booking, GBC traces backward to find that a summarizer agent dropped a key detail and rewrites its instructions.
Heard on the show
“So the method is called Gradient-Based Connections, and the core reframing is almost simple.”Episode 181 — How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires