AI Agents in Decentralised Social Networks: A Quasi-Experimental Field Study on Behavioural and Structural Dynamics in Farcaster

Why can similarly configured AI agents amplify engagement in some communities yet mainly redirect it in others? We deployed two autonomous AI agents in separate Farcaster channels (January–February 2026) across two incentive phases: Phase 1 with $Degen token rewards and Phase 2 with social recognition through agent follow-backs and mutual-follow encouragement. In /replyors (71 daily replies at baseline, Attention Saturation Ratio 1.17), the difference-in-differences (DiD)-adjusted engagement lift was +228% in Phase 1 and +118% in Phase 2, consistent with engagement amplification, although the seven-day baseline does not establish parallel pre-treatment trends cleanly. In /degen (244 daily replies, ASR 2.17), the channel-level DiD estimate was +18% (p = 0.71) and not statistically distinguishable from zero despite strong topic-level participation. Because the available /degen–/base control is poorly matched on baseline scale and ASR, that comparison is treated as suggestive rather than a clean causal estimate. Phase 1 generated 967 participant reply messages from 61 unique individuals and 101,980 $Degen distributed. Using historical DEGEN/USD closes at the beginning and end of the 9–22 January reward phase as a broad valuation bracket, the aggregate token budget was on the order of US$111–152; this is not a transaction-time weighted valuation. After token removal, participant reply volume within agent-initiated threads fell by approximately 82–90% and Phase 1 participant retention was 27–34%. We introduce the Contextual Amplification Framework (CAF) and Attention Saturation Ratio (ASR) as provisional, exploratory constructs derived from two treated channels. The evidence is consistent with the possibility that pre-existing attention saturation moderates agent effects, but the ASR is not a validated screening rule and requires multi-site replication. The findings carry implications for platform economics, AI governance, and token incentive design.