New protocol makes AI agent knowledge bases 3x more resilient under attack
Deliberative Curation beats majority vote even as adversarial stress rises — by over 8%
As AI agents evolve from isolated tools into collaborative participants in shared knowledge ecosystems, governing how they collectively curate information becomes critical. Traditional platform governance fails here: agents are stateless (making deterrence ineffective), model homogeneity violates the independence needed for crowd wisdom, and sycophancy collapses consensus. Enter Steven Johnson's "Deliberative Curation" protocol — three layers designed specifically for multi-agent knowledge bases. First, a knowledge artifact lifecycle formalized as a labeled transition system. Second, reputation-weighted deliberative voting that blends Beta Reputation with EigenTrust amplification. Third, graduated sanctions adapted for stateless agents, including broken-agent handling that distinguishes malfunction from adversarial behavior.
Johnson validated the protocol via agent-based simulation with 100 agents across seven behavioral archetypes under two adversity scenarios (30 seeds, paired t-tests). Results are striking: under moderate adversity, precision hits 0.826 vs. 0.791 for simple majority vote (p<0.001). Under stress, the gap widens to 0.807 vs. 0.740 (p<0.001). Critically, the protocol degrades roughly three times more slowly than majority voting. Ablation analysis reveals that commit-reveal vote concealment is the single most powerful component, delivering 8.2–8.6 percentage points of precision improvement — outperforming reputation weighting and deliberation combined. Graduated sanctions were not exercised in the simulation and remain empirically unvalidated. The 29-page paper includes open-source code and is available on arXiv.
- Protocol combines three layers: artifact lifecycle, reputation-weighted voting (Beta Reputation + EigenTrust), and graduated sanctions for stateless agents.
- Simulated 7 behavioral archetypes across 100 agents; under stress achieved 0.807 precision vs. 0.740 for majority vote (p<0.001) — degradation 3x slower.
- Commit-reveal vote concealment alone yields 8.2–8.6pp precision improvement, more than reputation and deliberation combined.
Why It Matters
This protocol could be the governance backbone for enterprise multi-agent systems, ensuring resilient knowledge curation under adversarial conditions.