Five agents that each hallucinate 5% of the time don't average out — they compound

Building a Multi-Agent System from Scratch: Architecture Patterns

Five agents that each hallucinate 5% of the time don't average out — they compound

A hands-on guide to splitting work across multiple agents, covering three architecture patterns: orchestrator-worker, blackboard, and pipeline. The author explains when each fits, then details the failure modes that only surface past a demo — infinite delegation loops, runaway token costs, conflicting concurrent writes, and hallucinations that downstream agents trust and amplify. The fix isn't better prompting but hard limits in code: delegation-depth caps, per-task budgets, and a real human-escalation path.

In a single-agent system, a hallucination is a wrong answer. In a multi-agent system, a hallucination is an input that gets compounded by every downstream agent that trusts it.

More from this day

2026-11-05