CPS Agent Architecture

Core Problem

• Converting CPS (Communication Proxy System) from chatbot to agent paradigm creates simultaneous design challenges
• Feature-listing approach (escalation, Q&A, monitoring, cross-question reasoning) makes inter-feature relationships unclear
• Agent needs to query human experts, but current escalation is "queue and wait" — inefficient
• Multi-turn conversations, approval workflows, multi-channel delivery (Slack/email/web) make Agent logic complex

Solution: 2 Channels + 1 Shared Capability + 1 Infrastructure Layer

1. Expert Channel (Bidirectional) — A2H Protocol

Abstract human experts as MCP tools: Agent-to-Human like Agent-to-Agent.

• Humans modeled as "slow tools" → orchestration logic simplified
• Context-rich escalation instead of bare handoffs

// A2H escalation with self-awareness
{
  "tool": "ask_expert",
  "context": "I know X but don't know Y.",
  "questions": [
    "(1) What is the timeline?",
    "(2) What is the current status?"
  ]
}

Key shift: queuing → conversation, self-awareness (knowing what you don't know).

2. User Channel (Unidirectional) — Q&A

• Dedicated Q&A responses
• Mode switches based on who's talking (Admin mode vs Q&A mode), not response depth adjustment

3. Messenger Layer as Tension Absorber

Agent decides "what to say", Messenger handles "how to deliver".

• One layer absorbs multiple concerns simultaneously: multi-turn management, approval workflows, multi-channel delivery

"Which layer should absorb this tension?" is THE key architectural framing.

4. Approval as Delivery Policy

• Approval is NOT an Agent tool — it's Messenger's delivery policy
• Agent doesn't need to know about approvals — just generates messages

5. Instance Separation as Access Control

• Instead of complex access_level schemas, separate CPS instances by domain
• Knowledge isolation achieved structurally

"Don't try to control — just separate." — Occam's razor at architecture level.

User Scenarios

Expert Channel

User asks about server migration:

1. Agent has no info → A2H tool call to expert with pinpoint questions
2. Expert responds asynchronously
3. Response goes through atomic decomposition → stored as memory
4. User gets answer

Approval Workflow

1. Agent generates answer
2. Messenger checks delivery policy → "this channel needs approval"
3. Admin approves → message sent

Agent is completely unaware of this process. Separation of concerns in action.

Multi-domain

• IT instance and HR instance run independently
• Natural knowledge isolation without access control logic
• Shared Messenger layer handles delivery for all instances