How it works

Five layers, one platform

Every request travels the same path through the YAAIF stack: from the channel it arrived on, through the agent that owns it, down to the model that reasons about it. Each layer has its own role, its own governance, and its own audit trail.

L1

Channels

Where work comes in

Microsoft Teams, web ChatUI, Claude Desktop, ChatGPT, email, voice, APIs, event streams, and any MCP-compatible client. Channel input is normalized into one contract before it reaches the agent.
L2

Agent

Goal-driven orchestration

The agent owns the request. It routes channel input, selects the execution path, and coordinates downstream steps. Policy gates and human approvals sit in line, not bolted on afterwards.
L3

Skills

Playbooks and routing

Skills map intent to reusable capabilities. They decide which tools should run for the current request and in what order, with the same playbooks reused across agents and channels.
L4

Capabilities

Execution against your systems

The execution layer invokes MCP tools and your existing integrations to do the actual work: writes to systems of record, queries against your data, triggers in downstream workflows.
L5

LLM Models

The reasoning engine

Any commercial provider or a self-hosted open model, hot-swappable per agent and per skill. Model choice is a platform decision, not a vendor lock-in.

Stylized representation of the YAAIF Control Center showing five layers: Channels (Teams, ChatUI, Claude Desktop, ChatGPT, plus a 'More' tile indicating Slack, email, voice, APIs, events and other extensible channels), Agent (orchestration), Skills (playbook and routing), Capabilities (execution), and LLM Models (reasoning engine). — Representative illustration of the YAAIF Control Center. Real product screens are not shown.

Layers are not silos. A single request crosses all five layers, and the Third Eye observability layer correlates the path into a structured trace. See the observability page for what that looks like in practice.