Chapter 31: Builder and Agent Roles

Last Updated: 2026-04

31.2 The Builder

Peter Westerman

Peter operates across the full scope of the work:

Domain	Responsibilities
Product	Product vision, user story definition, prioritization, acceptance testing, release decisions
Technology	Architecture decisions, infrastructure management, security oversight
Delivery	Build session management, agent invocation, shared library stewardship
Knowledge	Documentation curation, CLAUDE.md upkeep, knowledge base management

Peter’s primary interfaces:

• Cursor — primary development environment with AI agent access (development lane)
• Claude Code — long-horizon context management, CLAUDE.md maintenance (context lane)
• Antigravity — autonomous test dispatch, parallel debugging, browser QA (test/debug lane)
• n8n UI — workflow monitoring and management
• Dify console — knowledge base management
• Docker — infrastructure management

---

31.3 Specialist Agents — Peter’s AI Development Toolbox

The ITI Agent System is a catalog of specialist AI agents Peter invokes while building. They provide on-demand expertise in disciplines a solo builder would otherwise need to hire for. They are tools Peter directs; they do not make decisions, own outcomes, or operate without review.

The Orchestrator (start here)

The Orchestrator is the first agent Peter consults for any development question. It routes work to the correct specialist, coordinates multi-agent tasks, and knows about the shared library and all other agents. Think of it as the front door to the toolbox.

When to use: Any task that could involve multiple agents or where the right specialist is not obvious.

Specialist Agents

Agent	Specialty	Invocation
Pattern Agent	Architecture and design guidance; recommends shared library patterns	“What’s the best architecture for [feature]?”
API Integration Agent	Claude, Tavily, Pinecone, and third-party API integration	“Help me integrate [API] into [product]”
Database Agent	Schema design, migrations, query optimization	“Design the database schema for [feature]”
Template Agent	Project scaffolding from shared templates	“Scaffold a new WordPress plugin for [product]”
Migration Agent	Upgrades, refactors, data migrations	“Help me upgrade [product] from version X to Y”
Integration Agent	Cross-system integration (n8n, Dify, WordPress)	“Wire [product] to the n8n backend”
QA Agent	Test design, test writing, quality assurance	“Write tests for [feature]”
Documentation Agent	READMEs, CLAUDE.md files, inline docs	“Document this module”

---

31.4 Governance Agents

These agents help Peter enforce process and quality standards on his own work:

Agent	Role	When to Invoke
Vibe Coding Guardian	Audits builds against 15 pitfall categories	“Run a vibe coding guardrail audit on [product]”
Context Keeper	Maintains CLAUDE.md accuracy; enforces session protocol	“Check context drift on [product]”
Scope Owner	Evaluates scope changes; maintains parking lot	“Evaluate this scope change: [description]”
Claims Ombudsman	Audits documents for false or misleading claims	“Run a claims integrity audit on [document]”
Claims Evidence Curator	Maintains evidence registries; staleness checks	“Run a staleness check on claims evidence”

---

---

31.6 Working with the Agent Toolbox Effectively

Give agents the right context

Before invoking an agent, ensure it has access to:

• The project’s CLAUDE.md
• The relevant REQUIREMENTS.md section
• The ARCHITECTURE.md (for structural decisions)

One agent, one task

Agents perform best when given a single, well-defined task. Avoid mega-prompts that ask for multiple things at once. Break complex work into sequential steps, using the Orchestrator to coordinate.

Verify agent outputs

AI-generated code must be reviewed before committing. In particular, verify:

• Security rules are followed (Chapter 26)
• Shared library components are used where applicable (Chapter 13)
• The code matches the requirement in REQUIREMENTS.md
• Tests are written or updated

Agent conversation continuity

Within a single Cursor session, the AI retains context of the conversation. For long sessions that resume later, re-read CLAUDE.md at the start of the new session to re-establish context.

---

31.7 The Full Agent Inventory

The complete inventory of ~213 agents across all domains is in:

• ITI/operations/Agents/AGENTS-INDEX.md — human-readable index
• ITI/operations/Agents/AGENTS-INDEX.json — machine-readable index
• ITI/operations/agents-and-skills.md — global master reference

Before creating any new agent, check these indexes. Extending an existing agent is almost always better than creating a duplicate.

---

Previous: Chapter 30 — MCP Integrations | Next: Chapter 32 — Product Portfolio