Definition
Plain language
A rule that forces an AI research agent to pitch at least one bold, risky idea each round instead of only safe little tweaks.
As stated in the literature
In AutoScientists, a constraint on analyst agents requiring every proposal cycle to include at least one high-variance 'bold' proposal, with a public justification demanded when none exists; a hedge against over-exploiting incremental directions.
Why it matters: Without it, automated research agents tend to chase safe, small improvements and never explore the riskier ideas that lead to real breakthroughs.
For example, if an agent keeps suggesting tiny tweaks to a learning rate, the rule forces it to also propose trying a completely different model architecture, even if that might fail.
Heard on the show
“The analyst agents have what the paper calls an ambition quota.”Episode 095 — Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search