06 · Multi-Agent System

What you’ll learn

Instead of one agent doing everything, a multi-agent system gives each agent a narrow role and chains them together. Smaller, focused prompts are easier to control, test, and debug.

Our pipeline uses three roles:

• Researcher — gathers supporting evidence via retrieval.
• Analyst — interprets the evidence (and may call a tool).
• Writer — composes the final answer from evidence + analysis.

The trace shows each role’s contribution and the handoffs between them.

The handoff chain

flowchart LR
  R[Researcher<br/>retrieve evidence] --> A[Analyst<br/>summarise / use tool]
  A --> W[Writer<br/>compose answer]

import { multiAgentPipeline, loadCorpus } from '@lib/js';

const corpus = await loadCorpus('knowledge-base', import.meta.env.BASE_URL);
const { answer, trace } = multiAgentPipeline('Why do agents use tools?', corpus);

// trace actors: researcher → analyst → writer

from _shared.data import load_corpus
from multi_agent import multi_agent_pipeline

corpus = load_corpus("knowledge-base")
result = multi_agent_pipeline("Why do agents use tools?", corpus)
print(result["answer"])

Why specialise

• Simpler prompts. Each agent has one job, so its instructions stay short and reliable.
• Reusability. The researcher can serve many workflows.
• Debuggability. When output is wrong, the trace tells you which agent to fix.

A fixed chain works for predictable tasks. When you need to route work and run agents in parallel teams, you add a supervisor — that’s Agent Teams, next.