The Consensus Server Pattern: How to Catch AI Confabulation Before It Reaches Your Users

This article introduces the Consensus Server pattern to address the issue of AI confabulation, where large language models (LLMs) generate plausible-sounding but potentially inaccurate information. The Consensus Server uses multiple agents with distinct roles to independently evaluate findings and reach a consensus before surfacing information to users.

Details

The article explains that LLMs can confidently provide incorrect information, such as claiming a commit was added or an API endpoint returns a certain status code, even if those facts are wrong. This is a feature of how LLMs work, generating plausible-sounding text rather than verified facts. The author calls this 'confabulation' and argues that it can damage trust, break integrations, and send users down blind alleys in production AI systems. The classic solution of adding validation is insufficient, as you can't validate against a ground truth for every finding. The Consensus Server pattern addresses this by running multiple agents (Scout, Auditor, Dev) with distinct roles to independently evaluate and vote on each finding. The votes are weighted by the agent's confidence and aggregated to reach a consensus score. If the score clears a threshold, the finding is confirmed; otherwise, it's flagged for human review. The key insight is that an agent can be highly confident but wrong, so consensus catches disagreements that raw confidence scores hide. This makes consensus a more robust approach than validation alone.