Past a few thousand contributors, the human review systems that once felt fair quietly begin to fail. For open source projects at scale, selection committees burn out, and unconscious exclusion moves in. Mentorship slots go to the loudest voices. Months later, CNCF DevStats shows the damage.

24 months, one deployed stack, three FOSS communities with more than 15,000 contributors across APAC.

The talk presents an agentic contributor-operations stack we built on vendor-neutral primitives.

MCP servers wrapping GitHub, Slack, CHAOSS, and DevStats; single-purpose agents with AGENTS.md defined scope, gated by an audit layer with human-in-the-loop checkpoints.

Through the architecture walkthrough, attendees will see,

1. What the agents were and weren't allowed to do
2. Three production failures, told plainly, and how we recovered
3. Bias-prevention protocols that survived production

Leaving with a tested playbook, a deployment roadmap, and a clear sense of the way ahead.