Agents Party Compendium: Agent Orchestration Guide
The Compendium
The Agents Party is an interactive web tool that lets you visually configure and export Agent Orchestration packages. Using a fantasy RPG inspired theme, you can assemble a "Party" of specialized AI agents, customize their personalities, and download a ready-to-use orchestration setup tailored for platforms like Gemini or Claude. All entirely in your browser.
The Multi-Tier Architecture
The Agents Party framework uses a structured 3-tier hierarchy: The Throne Room (Meta-Level) for high-level strategy, The GM Session (Coordinator) for autonomous project management, and The Party (Sub-Agents) for specialized technical execution.
from actual coding. This ensures privacy, reduces clutter, and maintains clear roles.
I. The Throne Room (Meta-Level)
The strategic layer where you interact with meta-agents to define the quest, set constraints, and adjust rules mid-flight. It runs in an isolated directory to keep your main codebase clean.
- The Game Creator: The architect and your primary brainstorming partner.
- The Master of Spies: The real-time analyst, critiquing orchestration inefficiencies.
- The Bard: The chronicler who translates technical achievements into an engaging narrative recap.
II. The GM Session (Coordinator)
The Game Master (GM) runs inside the actual codebase. It acts as the project manager—reading journals, evaluating results, and spawning the correct Party Members to do the work. The GM coordinates but never writes code or runs tests itself.
Standard Quest Lifecycle
Phase 1: Recon
Phase 2: Execution (Iterative Loop)
Phase 3: Document & Report
III. The Party (Sub-Agents)
The specialized workers executing the tasks. The magic lies in their constructive tension: they are not yes-men, but experts who push back on bad plans and scrutinize each other's work.
IV. The Agentic Mindset (Best Practices)
The default presets are a culmination of custom ideas, ongoing trends in the broader agent ecosystem, and recent additions inspired by Simon Willison's Agentic Engineering Patterns. The system enforces strict behavioral guidelines to prevent "vibe coding" and AI hallucinations:
- Red/Green TDD: The Warrior is strictly instructed to write a failing test first before implementing any code. The Warlock will reject pull requests without tests.
- Context Minimization: The GM isolates sub-agents, passing them only the specific chunk of the plan they need to know, preventing prompt confusion.
- Aggressive Branching: Because writing code is cheap, the Party is encouraged to spin up multiple git branches to test different approaches simultaneously.
⚠️ Experimental Gotchas
To fully unlock the "true" parallel multi-agent potential of these
tools, you need to enable their experimental flags:
- Gemini (Subagents): Set "experimental": { "enableAgents": true } in your global settings to natively route to our .gemini/agents/ files.
- Codex (Multi-Agent): Ensure multi_agent = true is toggled via /experimental so it reads our .codex/config.toml.
- Claude (Agent Teams): Simply run the quest with the --team flag to switch
from hub-and-spoke subagents to a fully collaborative Agent Team.