Seven architecture principles for customer-facing agents, from a year in production

Anxious_Set2262 · reddit · 2026-11-04

A team that has spent over a year building agents that build customer-facing agents shares the architectural calls behind their current design:

The model proposes, code decides: identity, permissions, money, and allowed actions are enforced in code.
One agent with written procedures is the default; multi-agent teams only with good reason.
No single retrieval method fits: tool calls for live data, RAG for written knowledge, agentic search for complex questions.
Guardrails that guess fail both ways — must-stop rules go in exact code.
Agents fail silently: evals run on real conversations, and fixes ship only after business-side approval.
Containment alone proves little: a Taobao field study found agent-only chats rated 0.858 lower (5-point scale) than comparable chats with a human.
Text often isn't the answer: generative UI built at the moment of the question.

The authors frame this as judgment, not proven research, and openly invite critique.

More from coding & agent