Architecture of Clients.ai AI agents
Clients.ai is a multi-agent system, not one prompt in a loop. One part decides. One part writes. One part guards what may be offered. One part delivers. Each has a bounded job, and all of it runs on configuration you own. This page describes the architecture at the level we publish; the internals stay proprietary.
Your configuration is the source of truth
Everything starts from what you define: your offer, instructions, tone, channels, schedules, and limits. Agents act only within that envelope — no component invents offers or commitments you didn’t authorize.
Separation of responsibilities
Conversation handling is decomposed into specialized parts: components that judge whether and when to act, components that write, components that guard what may be offered, and components that deliver. A writer doesn’t authorize; a guard doesn’t write. That separation is what keeps output both natural and inside your standards.
Channels as adapters
LinkedIn, Instagram, and email are integrated as channels behind a common core, via established messaging-integration infrastructure. The conversational intelligence is shared; each channel adapter respects that platform’s specifics.
Decision boundaries and human control
Hard gates sit between deciding and sending: agent on/off, reply-only mode versus proactive, daily limits, and automatic pause when nothing should be running. If a conversation belongs with the human owner, the system can alert you and stand down in that thread. And before any of it touches real leads, you can test the agent yourself and iterate until it meets your bar.
Observability and outcomes
Actions are recorded as history — what was scored, sent, and answered — and post metrics (reactions, comments, impressions) feed back into what the content agents do next. Outcome tracking is part of the loop, not an afterthought.
What we deliberately don't publish
Prompts, model routing, scoring formulas, anti-abuse mechanisms, and infrastructure details are internal. Publishing them would be a blueprint for gaming the system rather than useful knowledge — the boundary of this page is the boundary of what’s safe to share.