Course Content
LangGraph Agents
7 sections · 49 lessons
When should you use multi-agent systems instead of one strong agent with tools?
What you need to know
Good reasons to split
- Tool overload — with dozens of overlapping tools, selection accuracy drops. Grouping by domain (billing, orders, technical) gives each agent a short, clear list.
- Different roles — a researcher needs search tools and a cheap fast model; a writer needs no tools and a strong model; a critic needs a rubric.
- Different permissions — the agent that reads untrusted web pages should not be the one that can pay refunds.
- Parallelism — ten documents analysed at once instead of one after another.
- Independent review — a critic that never saw the author's reasoning is less likely to share its mistakes.
- Ownership — separate teams deploy and evaluate their own agent.
Weak reasons
- It sounds more advanced.
- The single prompt is messy — fix the prompt.
- "Specialists" without evidence that the generalist fails.
Current patterns
LangChain's docs describe four main patterns:
| Pattern | How it works |
|---|---|
| Subagents (supervisor) | A main agent calls other agents as tools and keeps control |
| Handoffs | A tool call switches which agent is active |
| Router | A classification step sends the request to one or more agents, then results are combined |
| Custom workflow | A hand-built LangGraph graph mixing fixed steps and agents |
The separate langgraph-supervisor library now recommends building the supervisor directly with tool calling instead.
A real-life example
A telecom company built a five-agent support system (triage, billing, technical, retention, summariser). On a 300-case eval, it solved 81% at Rs 3.10 per case. A single agent with 9 well-described tools solved 79% at Rs 0.90. The team kept one agent and moved only retention offers to a separate agent, because that agent needed access to discount tools that the main agent must never have. Final result: 82% at Rs 1.05.
Follow-up questions to expect
- "What is the biggest hidden cost?" — Context loss at every handoff, and debugging: failures sit between agents, where no single prompt is at fault.
- "How many tools is too many for one agent?" — There is no fixed number; measure tool-selection accuracy on your own evals as you add tools, and split when it drops.
- "Is a workflow with several LLM calls a multi-agent system?" — Not necessarily. Fixed steps with LLM calls are a workflow; agents choose their own next actions.