Council Mode & Agent Teams

Council Mode brings multiple AI models into a structured, round-based discussion — turning single-model prompting into a deliberative process where different perspectives compete and converge.

Council Mode Overview

Instead of asking one model and hoping for the best, Council Mode seats several models at the same table and lets them debate.

Aspect	Detail
Round-based	Each round gives every participant one turn to speak before the next round begins
Structured output	Responses are threaded so you can follow each model's argument across rounds
Consensus tracking	Sciorex highlights points of agreement and flags unresolved disagreements
Configurable rounds	Set the number of rounds (1 to 10) or let the discussion run until pass detection triggers

INFO

Council Mode is not a simple "ask three models the same question." Each participant sees the full conversation history, including what the other models said, and responds in context.

Multi-CLI Debate

Pit different AI backends against each other for the same problem.

• Claude Code vs Gemini CLI vs OpenAI Codex — or any combination of configured backends
• Each participant uses its native strengths: Claude for nuanced reasoning, Gemini for breadth, Codex for code generation
• Responses are displayed in parallel columns with a unified diff when outputs are code
• You can intervene at any round to steer the discussion, add constraints, or ask follow-up questions

TIP

Multi-CLI debate is particularly effective for architecture decisions where different models may favor different trade-offs. Let them argue it out, then pick the approach that survives scrutiny.

Custom Personas

Define exactly who sits at the council table.

Field	Purpose
name	Display name shown in the conversation thread
systemPrompt	Instructions that shape the persona's behavior and expertise
temperature	Controls response randomness (0.0 for deterministic, 1.0+ for creative)
description	A short label shown in the participant list (e.g., "Security Auditor")
model	Optionally pin a persona to a specific backend model
icon	Icon identifier for the persona
tags	Tags for categorization and filtering

• YAML import/export — save persona configurations as .yaml files for version control and sharing
• Persona libraries — maintain sets of personas grouped by use case (e.g., "Code Review Panel", "Research Committee")

name: Security Auditor
systemPrompt: |
  You are a senior application security engineer.
  Focus on OWASP Top 10 vulnerabilities, authentication
  flaws, and data exposure risks. Be specific and cite
  CWE identifiers where applicable.
temperature: 0.3
description: Security Auditor
model: claude-opus-4-6

Turn Management

Council Mode uses a structured turn system to keep discussions productive.

• Sequential turns — each participant speaks once per round in a fixed order
• Pass detection — if a participant has nothing new to add, it can pass; when all pass, the discussion ends
• Structured rounds — rounds are numbered and labeled, making it easy to reference earlier arguments
• Moderator mode — optionally designate one participant as moderator to summarize each round and propose next steps
• Time limits — set per-turn token limits to keep responses focused and prevent runaway generation

WARNING

Pass detection relies on the model explicitly indicating it has nothing to add. Some models are reluctant to pass. If discussions run too long, set a hard round limit as a backstop.

Agent Teams

Go beyond discussion into coordinated multi-agent execution.

Role Type	Example	Responsibility
Team Lead	Orchestrator agent	Breaks down tasks, assigns work, reviews results
Specialist	Code Writer, Tester, Reviewer	Executes assigned subtasks with domain-specific tools
Observer	Logging agent	Records decisions and rationale without participating

• Agents communicate through structured message passing, not free-form chat
• Each agent has its own tool permissions, context window, and memory scope
• The team lead can reassign tasks dynamically based on intermediate results
• Agent teams can be saved as templates and reused across projects

INFO

Agent Teams differ from Council Mode in a key way: Council Mode is about discussion and deliberation, while Agent Teams are about coordinated execution. Use Council Mode to decide what to build; use Agent Teams to build it.

Community Marketplace

Share and discover personas and agent team configurations.

• Export — publish your persona or team template to the community marketplace
• Import — browse, preview, and install community-created configurations with one click
• Ratings and reviews — community feedback helps surface the most effective configurations
• Version tracking — persona authors can publish updates; importers receive upgrade notifications
• Categories — browse by domain (security, DevOps, research, writing) or by team size

TIP

Start with community personas to see what works, then customize them for your specific needs. A well-tuned Security Auditor persona from the marketplace can save hours of prompt engineering.

Use Cases

Council Mode and Agent Teams are designed for decisions and tasks that benefit from multiple perspectives.

Use Case	Setup	Why It Works
Architecture decisions	3 personas: Backend Lead, Frontend Lead, DevOps Engineer	Each brings domain-specific constraints to the table
Code review	2 personas: Author's Advocate, Devil's Advocate	Structured adversarial review catches more issues
Research hypothesis evaluation	4 personas: Domain Expert, Statistician, Skeptic, Synthesizer	Prevents confirmation bias by forcing critical examination
Debugging	2 models: Claude (reasoning) + Codex (code search)	Combines deep reasoning with broad code navigation
Technical writing	3 personas: Subject Expert, Editor, Reader Advocate	Produces accurate content that is also clear and accessible

Related

• Agents — single-agent configuration and tool permissions
• Chat Interface — the underlying conversation system Council Mode builds on
• Community Marketplace — browse and share configurations