The Idea

A single AI agent trying to research, analyze, write, and review its own work is like one person doing every job in a company. It works for small tasks, but for complex projects, specialization and collaboration produce far better results.

Multi-Agent Compositions assign different Level 2 patterns to different agents. A researcher agent uses ReAct with RAG to gather information. An analyst uses Tree-of-Thoughts to explore implications. A writer uses Skeleton-of-Thought for structured output. A reviewer uses Chain-of-Verification for quality control. Then these agents are coordinated through defined communication patterns — just like a human team.

Component Patterns

Each agent is powered by Level 2 compositions:

ReAct RAG Patterns Skeleton-of-Thought Meta-Prompting Reflexion Self-Consistency

The magic is not in any single pattern but in how specialized agents coordinate — each bringing the right capability at the right time.

Four Ways Agents Can Work Together

Pipeline

Sequential Handoff

Each agent completes its work and passes the result to the next. Like an assembly line — clear handoff points, predictable flow.

Researcher → Analyst → Writer → Reviewer
Hierarchy

Supervisor Delegation

A supervisor agent decomposes the task and delegates to workers. It decides who to call, what to ask, and how to combine results.

Supervisor → [Coder | Researcher | Tester]
Debate

Adversarial Review

Two advocates argue different positions across multiple rounds. A judge evaluates the debate and renders a verdict. Diverse reasoning reduces blind spots.

Advocate A ↔ Advocate B → Judge
Ensemble

Parallel Voting

Multiple agents solve the same task independently using different patterns. Their answers are combined by voting or synthesis. Independence improves reliability.

[Agent 1 | Agent 2 | Agent 3] → Aggregator

Agent Roles

ReAct + RAG
Researcher

Gathers info from tools and knowledge bases

Tree-of-Thoughts
Analyst

Explores implications through branching reasoning

Skeleton-of-Thought
Writer

Produces structured, organized output

Verification
Reviewer

Checks facts, catches errors, ensures quality

Meta-Prompting
Supervisor

Orchestrates workers, decides delegation

Reflexion
Critic

Provides constructive feedback for improvement

See It in Action

A research report pipeline using the sequential topology.

1
Research Phase
Researcher using ReAct + RAG
Searched company database, industry reports, and recent news. Found 12 relevant sources on market trends, competitor analysis, and customer feedback. Compiled structured findings with citations.
↓ findings passed to analyst
2
Analysis Phase
Analyst using Tree-of-Thoughts
Explored three interpretations of the data: optimistic (market expansion), cautious (competitive pressure), and contrarian (niche opportunity). Recommended the cautious path with strongest evidence support.
↓ analysis passed to writer
3
Writing Phase
Writer using Skeleton-of-Thought
Generated outline first: Executive Summary, Market Analysis, Competitive Landscape, Recommendations, Risk Factors. Expanded each section with data from the analysis. Produced 8-page report.
↓ draft passed to reviewer
4
Review Phase
Reviewer using Chain-of-Verification
Verified 3 of 4 key claims against original sources. Flagged one unsupported statistic. Checked logical consistency of recommendations against evidence. Sent revision request for the flagged claim.

Why This Works

Specialization works for the same reason it works in human teams: each agent can be optimized for its specific task. A researcher agent's prompt is tuned for thoroughness and source finding. A reviewer's prompt is tuned for skepticism and error detection. These are fundamentally different cognitive modes.

The communication topology adds a second layer of value. In a pipeline, each stage refines the previous. In a debate, adversarial pressure catches blind spots neither side would find alone. In an ensemble, independent approaches provide robustness against any single method's weaknesses.

The System

Give each agent a specialized role and the right technique for that role. Connect them through pipelines, hierarchies, debates, or ensembles. The team is greater than the sum of its parts.

When to Use This

When to Skip This

How It Relates

Each agent in a multi-agent system can internally run a Cognitive Loop. For tool-heavy tasks, agents can delegate to JARVIS. For hard reasoning within a single agent, LATS can plug in. The Adaptive Pattern Router can choose which topology to use based on the task.

At Level 4, Federated Agent Networks extend this concept to distributed systems where agents collaborate without central control, and Collective Intelligence patterns emerge from large-scale agent interaction.