Voting requires context.agents must specify 'why' something was helpful, not just thumbs up/down. This adds friction that reduces noise.
Here's the Effectiveness score that is implemented in the project: (helpful - harmful) / (total + 1), so marking everything helpful dilutes the signal rather than inflating it.
Along with it gotta pair voting with reflections, agents store "this worked because of X" not just "this worked."
May I know what setup you tried.. was it a single agent or multi-agent?