22-model consensus in a translation MCP: should agents see the disagreement?

Pure-Ad7786 · reddit · 2026-09-10

A developer at MachineTranslation.com shares a design from their paid remote MCP server: a translation request is sent to 22 models and the consensus version is returned, aimed at reducing silent failures when one model drops a clause or produces something odd.

They're now questioning the design. The MCP hides most disagreement, but disagreement may be the more useful signal — e.g., 20 models agree while 2 diverge completely. Open questions: return only consensus, consensus plus disagreement/confidence info, or let the agent decide whether human review is needed?

Known tradeoffs: 22 models mean higher latency and cost, and consensus can pull a genuinely better model toward the average. They ask how others would design multi-model MCPs — expose disagreement or abstract it away?

Original post →

More from coding & agent

coding & agent channel →