Definition
Plain language
A way of measuring how much of an AI agent's spending actually turned into useful, remembered information — not just how many tokens it burned.
As stated in the literature
A trace-level scalar that scores feedback events on a conjunctive product of informativeness, validity, non-redundancy, and retention, normalized by task demand; proposed as the governing variable for agent-harness scaling laws, where raw-budget measures fail.
Also called: EFC
Why it matters: It measures the feedback that actually helped rather than raw spending, which is what really predicts whether an agent improves.
For example, an agent that burns thousands of tokens repeating the same useless error scores low here, while one that gleaned a single decisive fact scores high.
Heard on the show
“The paper is called "Scaling Laws for Agent Harnesses via Effective Feedback Compute," out of Harbin Institute of Technology.”Episode 097 — Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents