Definition
Plain language
An experiment where researchers paste a clarifying message into an agent's conversation at a specific moment, instead of waiting for the agent to ask.
As stated in the literature
A methodology for measuring value-of-information curves in long-horizon agents by deterministically injecting a synthetic user message at a calibrated trajectory position and measuring downstream success.
Why it matters: Controlled timing of clarifying information lets researchers measure exactly when help is worth giving, which informs when agents should ask for it.
For example, halfway through an agent's task, researchers paste in a synthetic user message saying 'use the staging database' and then measure whether the final result improves.
Heard on the show
“They call it forced injection.”Episode 035 — Why Frontier Agents Ask for Clarification at Exactly the Wrong Moment