Building a Council of AI Agents + Designing a Persona System

The Problem

A single AI assistant is a generalist — competent across many domains but expert in none. For tasks requiring specialized knowledge (security auditing, performance optimization, architectural review), you need multiple specialized perspectives. The challenge: how do you design a system where multiple AI agents collaborate effectively?

The Solution

The Council System

I built a 4-member council system where each member has a specialized role:

Member	Role	Expertise
Architect	System design review	Architecture patterns, trade-offs, scalability
Coder/Debugger	Code quality & bug diagnosis	Error patterns, edge cases, fix recommendations
Security	Vulnerability assessment	Threat modeling, secure coding, OWASP
Performance	Optimization review	Bottlenecks, caching, algorithms

A dedicated @council trigger in AGENTS.md activates the entire council for any task.

Persona System (Maestro)

Beyond the council, I designed a Maestro persona — an orchestration layer that coordinates AI agents like a conductor:

• SOUL.md — The persona definition file containing Maestro's identity, skills, and operating principles
• Persona section added to AGENTS.md to load Maestro on session start
• Decision framework — Maestro decides which council members to consult for each task

The persona system follows this architecture:

Maestro (Orchestrator)
  ├── Council: Architect
  ├── Council: Coder/Debugger
  ├── Council: Security
  ├── Council: Performance
  └── Direct: html-reports, brainstorming, critique skills

Subagent → Council Migration

Three existing subagents were migrated into the council structure:

• executing-plans/ → deleted (replaced by council workflow)
• subagent-driven-development/ → deleted
• writing-plans/ → deleted
• council/SKILL.md → created with 4 members

Key Results

• 4-member council operational — each member with distinct expertise
• @council trigger works from any context in AGENTS.md
• Maestro persona created with full identity, skills, and principles
• 3 old subagent systems cleaned up and replaced
• Test protocol documented: restart OpenCode and verify 0mem auto-load + Maestro persona

Key Decisions

Decision	Rationale
Council over single agent	Multiple perspectives catch more issues
Maestro as orchestrator	Separates "what to do" from "how to do it"
@council trigger	Simple, memorable invocation from any context
SOUL.md for persona	Separate file keeps persona distinct from config
Delete old subagents	Prevents confusion with the new council system

Takeaways

1. Multi-agent systems need an orchestrator. Without Maestro, the council has no direction.
2. Specialization beats generalization. Four focused reviewers outperform one generalist.
3. Personas give AI consistent behavior. A well-defined persona file shapes every response.
4. Clean up old patterns. Migrating from subagents to council left no obsolete files behind.