One popular solution lately has been instead of banning too much, because of the danger of false positives, to use vouch [0]. Trusted people get vouched and you prioritize their actions. Unknown people (or agents) need to gain trust to be vouched and bad actors can still be banned.

[0]: https://github.com/mitchellh/vouch