Agent Orchestration System
Relevant source files
The following files were used as context for generating this wiki page:
The Agent Orchestration System is a modular framework designed to delegate specialized domain tasks to autonomous subagents within the Everything Claude Code (ECC) ecosystem. By shifting from a monolithic "single-model" approach to a multi-agent architecture, the system optimizes token economics, enhances accuracy through specialized knowledge, and enables parallel execution of complex software engineering workflows.
The system utilizes an Agent-First philosophy, where the primary orchestrator (Claude Code) proactively identifies requirements and dispatches specialized agents—such as planners, architects, and reviewers—to handle discrete phases of the development lifecycle. This orchestration is governed by specific model policies, defined triggers, and standardized tool invocation protocols to ensure consistent, production-ready output.
Sources: AGENTS.md:7-12, the-longform-guide.md:204-209, CLAUDE.md:1-12
Core Architecture and Multi-Model Collaboration
The orchestration system is built on a tiered architecture where the main Claude instance acts as the central Orchestrator. It manages state, directs traffic, and synthesizes outputs from specialized models tailored to specific tasks.
Model Selection and Specialization
The system employs different models based on the reasoning depth required for a task:
- Haiku: Used for low-sensitivity tasks like single-file edits, documentation, and simple exploration/search to reduce costs.
- Sonnet: The default for 90% of coding tasks, balancing performance and context nuance.
- Opus: Reserved for complex architectural decisions, deep reasoning, and security-critical analysis.
Sources: the-longform-guide.md:126-150, CONTRIBUTING.md:168-174
Orchestration Data Flow
The following diagram illustrates the lifecycle of a request through the orchestration system, specifically during a collaborative development workflow.
This diagram shows the sequential phases (Research → Ideation → Plan → Execute → Optimize → Review) used in collaborative workflows.
Sources: commands/multi-workflow.md:124-177, the-longform-guide.md:218-226
Specialized Agent Registry
The system provides 67 specialized agents, each with a defined purpose and specific trigger conditions. These agents are defined using Markdown files with YAML frontmatter at agents/*.md.
| Agent | Core Purpose | When to Invoke |
|---|---|---|
| planner | Implementation planning | Complex features, refactoring |
| architect | System design and scalability | Architectural decisions, new subsystems |
| code-reviewer | Quality and maintainability | Immediately after writing/modifying code |
| tdd-guide | Test-driven development | New features or bug fixes |
| security-reviewer | Vulnerability detection | Before commits, sensitive data handling |
| doc-updater | Documentation and codemaps | Proactively for code changes/API updates |
| loop-operator | Autonomous loop execution | Monitoring stalls, intervening in long runs |
Sources: AGENTS.md:19-50, agents/doc-updater.md:5-15, CLAUDE.md:22-31
Execution Workflows
Orchestration is carried out through two primary methods: Sequential Phasing and Parallel Dispatch.
Sequential Phase Pattern
The orchestrator evaluations sub-agent returns through iterative loops, often passing objective context across five specific phases:
- RESEARCH: Uses the Explore agent to generate a
research-summary.md. - PLAN: Uses the planner agent to create a
plan.md. - IMPLEMENT: Uses the tdd-guide agent for implementation.
- REVIEW: Uses the code-reviewer agent to generate
review-comments.md. - VERIFY: Uses the build-error-resolver to ensure stability.
Sources: the-longform-guide.md:218-226, commands/multi-workflow.md:144-177
Team Builder (Parallel Dispatch)
The Team Builder skill allows for the interactive composition of agent teams. This mechanism spawns multiple agents in parallel to perform independent analysis on the same task.
This sequence shows how the system manages independent agent calls simultaneously without inter-agent communication.
Sources: skills/team-builder/references/summary-1.md:1-25
Implementation Standards for Agents
Every agent within the orchestration system must adhere to a standardized format to ensure compatibility with the Orchestrator's dispatch tools.
Agent Configuration Format
Agents are configured via frontmatter that defines their identity and tool access:
---
name: code-architect
description: Designs feature architectures by analyzing patterns...
model: sonnet
tools: [Read, Grep, Glob, Bash]
---
Sources: agents/code-architect.md:1-6, RULES.md:19-24
Orchestration Guidelines
- Proactive Delegation: The system is designed to use agents proactively without waiting for user prompts (e.g., calling
code-reviewerimmediately after code is written). - Context Isolation: Sub-agents save context by returning summaries rather than dumping full logs into the orchestrator's window.
- Parallelization Boundaries: Parallel execution is used for independent operations, while sequential execution is mandatory when outputs from one phase are required as inputs for the next.
Sources: AGENTS.md:52-60, the-longform-guide.md:210-217, CLAUDE.md:56-62
Summary
The Agent Orchestration System provides a structured environment for multi-agent collaboration, enabling the Everyting Claude Code project to handle large-scale refactors and complex feature implementations. By utilizing specialized roles—from code-architect to security-reviewer—and enforcing strict phase-based workflows, the system ensures that every code change is planned, implemented via TDD, and reviewed for quality and security. This modular approach allows for scalable development where the orchestrator manages the high-level vision while specialist agents execute domain-specific logic.