elit3-x/elitex-skill-library/agent-architecture

Agent Architecture

Source: Knowledge Base frameworks — agent-responsibility-minimization, multi-agent-failure-patterns, phronesis-practical-wisdom

When to Use

Trigger this skill when:

• Designing any AI agent, automation, or agentic workflow
• Building with n8n, Make.com, Relevance AI, or any agent platform
• User mentions "agent", "automation", "workflow", "multi-agent"
• Architecting client AI systems for EliteX or Baukasten IT

Core Rule

Agent = Decision-making + Communication ONLY

An agent that does generation is an overloaded agent. Offload it.

Default: Single Agent

Google Research proved that adding agents makes systems worse, not better:

• Coordination overhead grows super-linearly
• Communication noise degrades signal across handoffs
• Optimal team size is smaller than intuition suggests
• A single well-configured agent frequently outperforms multi-agent swarms

Only add agents when tasks are truly independent with no shared state. The moment agents need to coordinate, the system degrades.

Always measure multi-agent performance against a single-agent baseline.

Sub-Agent Communication Template

For each tool or sub-agent in a system, define exactly three things:

### [Tool/Sub-Agent Name]
**What it does:** [One sentence]
**When to use it:** [Trigger condition]
**How to communicate:** [Example message with variables]
 
Example:
  Input: "Research prospect {{company_name}} and return company size, industry, and recent news"
  Output: JSON with company_size, industry, recent_news fields

Phronesis: Principles Over Rules

Write principles-based guidance for agents, not rigid rule sets:

• A phronesis-oriented agent understands WHY rules exist
• It weighs context against principles and makes judgment calls
• Rigid rules break on edge cases; principles adapt

Bad: "Never send emails after 5pm" Good: "Respect the recipient's working hours. Consider timezone and urgency."

Architecture Patterns

Pattern 1: Micro-Agent Per Integration

Build one agent per software integration (email agent, calendar agent, CRM agent). Reusable across projects.

Pattern 2: Deterministic Shell, Agentic Core

Keep deterministic steps as deterministic logic (templates, fixed rules, conditional branches). Use agents ONLY for parts requiring dynamic reasoning.

Pattern 3: Human-in-the-Loop Checkpoints

AI generates → human reviews/edits → AI continues. Captures user decisions as feedback for self-improvement.

Discovery Method: Dream Day Interview

When scoping agent work for clients, ask:

"Describe when you do your job perfectly vs. an average day."

The gap reveals 2-3 automatable workflows — the best starting point for any agentic build.

Anti-Patterns

• Monolithic agent: One agent doing everything (generation + decisions + communication)
• Agent sprawl: Adding agents without proving independent tasks exist
• Rigid SOPs: Step-by-step procedures that break on edge cases
• No baseline: Building multi-agent without testing single-agent first
• Coordination theater: Agents passing messages back and forth with no value added