Make Agent Fast

Context

Make Agent Fast is a company-facing platform for launching customer agents on the website a business already has and in the messaging channels its customers already use. Each agent has an explicit job: general assistance, sales and enquiries, customer service, or booking.

Problem

A useful customer agent needs more than an open-ended chat box. It has to answer from approved company material, know when it cannot confirm something, collect a structured request, hand the conversation to a person, and behave consistently across channels. The setup also has to be manageable by a company owner without turning agent configuration into a software project.

Approach

I designed the product around a review-before-publish workflow. A company can paste its website, let the platform crawl and organize the material, review the drafted agent, choose its actions and languages, and then publish it through an embed or messaging connector. The runtime keeps knowledge, permissions, metering, conversations, and handoff boundaries behind one agent identity.

What Shipped

• Website crawling, source ingestion, pgvector retrieval, and approved-knowledge answers.
• One embeddable customer widget with configurable position, label, theme, language, text, voice messages, and live voice on eligible plans.
• Telegram, WhatsApp, Messenger, Instagram, Discord, and KakaoTalk connectors sharing the same knowledge and owner inbox.
• General, sales, service, and booking roles with request capture and explicit human handoff.
• EN, KO, and UZ platform interfaces, with agents able to speak EN, KO, UZ, and RU.
• Polar subscriptions and credit metering, consent-aware voice features, audit records, deployment checks, and production failover controls.

Results

• The product is live at makeagent.fast with a card-free trial path for launching a customer agent.
• Companies can start from their existing website instead of writing an agent prompt and knowledge base from scratch.
• The same agent can answer, capture, book, or hand off without separating each channel into a different support system.

Technical Notes

The platform is a Next.js 16 and React 19 application backed by Drizzle ORM, PostgreSQL, and pgvector, with PGlite for local development and tests. Its provider layer supports metered language and voice services, retrieval keeps provider and model identity attached to every chunk, and credit reservations happen before external calls. The customer widget is served as an embed loader and isolated iframe, while messaging connectors route conversations into the same product model.