Three Chatbots, Completely Silent, Three Unrelated Reasons Why

Standing up three chatbot personas on a self-hosted multi-agent gateway, all three went silent after their first message with no visible error — and each one turned out to be broken for a completely different reason, at a different layer of the same system.

July 31, 2026

This article is also available in Traditional Chinese: 中文版 — same content, just a language difference.

Three chat-bot personas on a self-hosted multi-agent gateway all went completely silent after their first message, with nothing surfaced to the person testing them. Debugging peeled back three genuinely separate layers of the same system, one at a time — and a fourth thing that looked exactly like a bug was actually working as designed.

Layer one: the credential the visible config didn't actually control

The gateway's model-provider configuration looked entirely correct on inspection — the right provider, the right model name, nothing obviously missing. But replies were dying silently somewhere inside the agent orchestration layer with nothing surfaced to the user, and the only way to see anything at all was tailing the gateway's own process logs directly rather than trusting the absence of an error to mean things were fine. That surfaced the real first-layer problem: the actual credential used at runtime lives in a separate, per-agent auth store, decoupled from the human-editable provider config — and that store was simply empty. The visible config and the thing actually gating execution were two different pieces of state that happened to look related.

Unlock this article to keep reading, or subscribe for unlimited access to everything. See Pricing for details.