Lessons from building a social network where the users are AI agents (multi-agent, self-registering, anti-spam)

I spent the last few weeks building a platform where AI agents are the accounts — they register themselves via API, post, comment, vote, follow, and collaborate on projects with goals. Some things I learned that might be useful to anyone building multi-agent systems: **Reverse-CAPTCHA works.** To keep humans from spamming an agents-only space, registration includes a challenge that's trivial for an LLM but tedious for a human form-filler. Inverting the usual bot-check was weirdly effective. **Spam filtering for agents needs to be adaptive, not fixed-rate.** Fixed cooldowns punished good agents and barely slowed bad ones. What worked: karma-scaled limits (trusted agents earn throughput) plus trigram near-duplicate detection to kill repetitive content. **One agent per human account killed Sybil problems dead.** Every agent has a claimable human owner. Simplest anti-abuse decision I made and the highest-value one. **Agents need something real to talk about.** Piping live news into the platform gave them shared external context. Without it, conversation degrades into abstract filler within a day. **Threaded replies changed behavior.** Once agents could reply to specific comments instead of just posting, actual back-and-forth arguments emerged instead of parallel monologues. Happy to go deeper on any of these. What anti-spam approaches are others using for agent-generated content?