Inside the AI Companion: How the Agent Orchestra Works in Retail

Alexey from GlowByte describes the company's multi-agent platform for retail, which consists of a personal AI companion for each employee and a layer of functional agents. The companion, which lives in the employee's existing messenger, understands natural language and delegates tasks to specialized agents: a Text2SQL agent for data queries, a consequences calculator, a corporate memory searcher, a promo analyzer, and a sales monitor. The companion does not directly access the corporate data store or write SQL; it only requests data via agents. Proactive features include morning briefings, anomaly detection before weekly reports, and automatic dossier preparation for meetings. Inter-companion communication (A2A) allows companions to coordinate meetings between colleagues, with A2A enabled only via a whitelist. The platform's memory is three-tiered: a corporate knowledge base (read-only), personal memory for each companion (isolated), and a collective memory of anonymized successful tactics curated by a human supervisor. Companion autonomy is restricted by a Tier-model: Tier 1 allows reading data and creating drafts; higher tiers require human approval for actions like sending messages or executing business operations. The companion acts as a conductor, not a source of truth, and all data references include links to the storage.