L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

retellai-architecture-variants

retellai-architecture-variants

Retell AI architecture variants — AI voice agent and phone call automation. Use when working with Retell AI for voice agents, phone calls, or telephony. Trigger with phrases like "retell architecture variants", "retellai-architecture-variants", "voice agent".

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: retellai-architecture-variants description: "Retell AI architecture variants \u2014 AI voice agent and phone call
\ automation.\nUse when working with Retell AI for voice agents, phone calls, or
\ telephony.\nTrigger with phrases like "retell architecture variants", "retellai-architecture-variants"
, "voice agent".\n" allowed-tools: Read, Write, Edit, Bash(npm:), Bash(curl:), Grep version: 1.9.0 license: MIT author: Jeremy Longshore jeremy@intentsolutions.io tags:

  • saas
  • retellai
  • voice
  • telephony
  • ai-agents compatibility: Designed for Claude Code

Retell AI Architecture Variants

Overview

Implementation patterns for Retell AI architecture variants — voice agent and telephony platform.

Prerequisites

  • Completed retellai-install-auth setup

Instructions

Step 1: SDK Pattern

import Retell from 'retell-sdk';
const retell = new Retell({ apiKey: process.env.RETELL_API_KEY! });

const agents = await retell.agent.list();
console.log(`Agents: ${agents.length}`);

Output

  • Retell AI integration for architecture variants

Error Handling

ErrorCauseSolution
401 UnauthorizedInvalid API keyCheck RETELL_API_KEY
429 Rate LimitedToo many requestsImplement backoff
400 Bad RequestInvalid parametersCheck API documentation

Examples

Choose a queue-backed architecture for bursty inbound calls

For a campaign that creates short spikes, keep the public number attached to a stable routing layer and place CRM enrichment behind an asynchronous queue. Start with a preview agent that records only synthetic test calls, set a clear timeout for the enrichment request, and configure a human-transfer fallback. Measure latency and transfer rate during the canary before selecting the variant for production traffic.

Resources

Next Steps

See related Retell AI skills for more workflows.

Skills enregistrées