Skip to content
kaleemsajid.com

// writing

Blog

Notes on building software well — architecture, reliability, and the non-code parts that make projects succeed.

· 9 min read

Building a multi-agent meeting that actually disagrees with you

A single model told to play three advisors just performs a debate with itself, and it always converges on a tidy consensus. Real dissent has to be enforced in the architecture. Here is how Camelot wraps a stateful, opinionated AI meeting around a deterministic core.